The Psychophysiology of Real-Time FinancialRisk Processing

Andrew W Lo1 and Dmitry V Repin2

Abstract

amp A longstanding controversy in economics and finance iswhether financial markets are governed by rational forces or byemotional responses We study the importance of emotion inthe decision-making process of professional securities tradersby measuring their physiological characteristics (eg skinconductance blood volume pulse etc) during live tradingsessions while simultaneously capturing real-time prices fromwhich market events can be detected In a sample of 10

traders we find statistically significant differences in meanelectrodermal responses during transient market eventsrelative to no-event control periods and statistically significantmean changes in cardiovascular variables during periods ofheightened market volatility relative to normal-volatility con-trol periods We also observe significant differences in thesephysiological responses across the 10 traders that may besystematically related to the tradersrsquo levels of experience amp

INTRODUCTION

The spectacular rise of US stock market prices in thetechnology sector over the past few years and the evenmore spectacular crash last year has intensified the well-worn controversy surrounding the rationality of invest-ors Most financial economists are advocates of thelsquolsquoEfficient Markets Hypothesisrsquorsquo (Samuelson 1965) inwhich prices are determined by the competitive tradingof many self-interested investors and such trading elim-inates any informational advantages that might existamong any members of the investment communityThe result is a market in which prices lsquolsquofully reflect allavailable informationrsquorsquo and are therefore unforecastable

Critics of the Efficient Markets Hypothesis argue thatinvestors are oftenmdashif not alwaysmdashirrational exhibitingpredictable and financially ruinous biases such as over-confidence (Barber amp Odean 2001 Gervais amp Odean2001 Fischoff amp Slovic 1980) overreaction (DeBond ampThaler 1986) loss aversion (Odean 1998 Shefrin ampStatman 1985 Kahneman amp Tversky 1979) herding(Huberman amp Regev 2001) psychological accounting(Tversky amp Kahneman 1981) miscalibration of proba-bilities (Lichtenstein Fischoff amp Phillips 1982) andregret (Clarke Krase amp Statman 1994 Bell 1982)The sources of these irrationalities are often attributedto psychological factorsmdashfear greed and other emo-tional responses to price fluctuations and dramaticchanges in an investorrsquos wealth Although no clearalternative to the Efficient Markets Hypothesis has yetemerged a growing number of economists psycholo-gists and financial-industry professionals have begun to

use the terms lsquolsquobehavioral economicsrsquorsquo and lsquolsquobehavioralfinancersquorsquo to differentiate themselves from the standardorthodoxy (Shefrin 2001) The fact that the currentvalue of the Nasdaq Composite Index a bellwetherindicator of the technology sector is 164634 (October17 2001)mdashonly 326 of its historical high of 504862(March 10 2000) reached less than 2 years agomdashlendscredence to the critics of market rationality Such criticsargue that either the earlier run-up in the technologysector was driven by unbridled greed and optimism orthat the precipitous drop in value of such a significantportion of the US economy must be due to irrationalfears and pessimism

However recent research in the cognitive sciencesand financial economics suggest an important link be-tween rationality in decision making and emotion (Loe-wenstein 2000 Peters amp Slovic 2000 Lo 1999 Elster1998 Damasio 1994 Grossberg amp Gutowski 1987)implying that the two notions are not antithetical butin fact complementary

In this study we attempt to verify this link experi-mentally by measuring the real-time psychophysiologicalcharacteristicsmdashskin conductance blood volume pulse(BVP) heart rate (HR) electromyographical signalsrespiration and body temperaturemdashof professional se-curities traders during live trading sessions Using port-able biofeedback equipment we are able to measurethese physiological characteristics in a traderrsquos naturalenvironment without disrupting their workflow whilesimultaneously capturing real-time financial pricing datafrom which market events can be defined By matchingthese events with the tradersrsquo psychophysiological re-sponses we are able to determine the relation between1MIT Sloan School of Management 2Boston University

copy 2002 Massachusetts Institu te of Technology Jou rnal of Cogn itive Neuroscience 143 pp 323- 339

financial risk measures and emotional states and dynam-ics In a pilot sample of 10 traders we find statisticallysignificant differences in mean electrodermal responsesduring transient market events as compared with no-event control baselines and statistically significant meanchanges in cardiovascular variables during periods ofheightened market volatility as compared with normal-volatility control baselines We also observe significantdifferences in mean physiological responses among the10 traders that may be systematically related to theamount of trading experience

In studying the link between emotion and rationaldecision making in the face of uncertainty financialsecurities traders are ideal subjects for several reasonsBecause the basic functions of securities trading involvefrequent decisions concerning riskreward trade-offstraders are almost continuously engaged in the activitythat we wish to study This allows us to conduct ourstudy in vivo and with minimal interference to andtherefore minimal contamination of the subjectsrsquo natu-ral motives and behavior Traders are typically providedwith significant economic incentives to avoid many ofthe biases that are often associated with irrational invest-ment practices Moreover they are highly paid profes-sionals that have undergone a variety of trainingexercises apprenticeships and trial periods throughwhich their skills have been finely honed Thereforethey are likely to be among the most rational decisionmakers in the general population hence ideal subjectsfor examining the role of emotion in rational decision-making processes Finally due to the real-time nature ofmost professional trading operations it is possible toconstruct accurate real-time records of the environmentin which traders make their decisions namely thefluctuations of market prices of the securities that theytrade With such real-time records we are able to matchmarket events such as periods of high price-volatilitywith synchronous real-time measurements of physiolog-ical characteristics

To measure the emotional responses of our subjectsduring their trading activities we focus on indirectmanifestations through the responses of the autonomicnervous system (ANS) (Cacioppo Tassinary amp Bernt2000) The ANS innervates the viscera and is responsiblefor the regulation of internal states that are mediated byinternal bodily as well as emotional and cognitiveprocesses ANS responses are relatively easy to measuresince many of them can be measured noninvasively fromexternal body sites without interfering with cognitivetasks performed by the subject ANS responses occur onthe scale of seconds which is essential for investigationof real-time risk processing

Previous studies have focused primarily on time-aver-aged (over hours or tens of minutes) levels of autonomicactivity as a function of task complexity or mental strainFor example some experiments have considered thelink between autonomic activity and driving conditions

and road familiarity for nonprofessional drivers (Brownamp Huffman 1972) and the stage of flight (take-offsteady flight landing) in a jetfighter flight simulator(Lindholm amp Cheatham 1983) Recently the focus ofresearch has started to shift towards finer temporalscales (seconds) of autonomic responses associated withcognitive and emotional processes Perhaps the mostinfluential set of experiments in this area was conductedin the broad context of an investigation of the role ofemotion in decision-making processes (Damasio 1994)In one of these experiments skin conductance re-sponses (SCRs) were measured in subjects involved ina gambling task (Bechara Damasio Tranel amp Damasio1997) The results indicated that the anticipation of themore risky outcomes led to more SCRs than of the lessrisky ones The brain circuitry involved in anticipatingmonetary rewards has also been localized (BreiterAharon Kahneman Dale amp Shizgal 2001) Anotherstudy (Frederikson et al 1998) reported neuroanatom-ical correlates of skin conductance activity with the brainregions that also support anticipation affect and cogni-tively or emotionally mediated motor preparation Re-cent experiments (Critchley Krase amp Statman 1999)support the significance of autonomic responses duringrisk-taking and reward-related behavior They providemore details on brain activation correlates of peripheralautonomic responses and also claim the possibility ofdiscriminating the activity patterns related to changingversus continuing the behavior based on the immediategainloss history in a gambling task

We focus on five types of physiological characteristicsin our study skin conductance cardiovascular data (BVPand HR) electromyographic (EMG) data respirationrate and body temperature

SCR is measured by the voltage drop between twoelectrodes placed on the skin surface a few centimetersapart Changes in skin conductance occur when eccrinesweat glands that are innervated by the sympathetic ANSfibers receive a signal from a certain part of the brainThree distinct lsquolsquobrain information systemsrsquorsquo can poten-tially elicit SCR signals (Boucsein 1992) The affectarousal system is capable of generating a phasic re-sponse (on the scale of seconds) that is associateddirectly with the sensory input indicating attentionfocusing or defense response and the amygdala is theprimary brain region involved The emotional responseto a novel or highly complex task is an example of affectarousal Another system that can initiate a phasic SCRcenters on the basal ganglia and is related to a prepar-atory activation mediated by internal cognitive pro-cesses The body exhibits increased perceptual andmotor readiness and high attention levels Expectationof an event or preparation to an important actionillustrates the operation of this system The third systemoften called the lsquolsquoeffort systemrsquorsquo produces tonic changesin the level of skin conductance and is related to thelong-term changes of a general emotional (hedonic)

324 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

state or attitude indicating a nonspecific increase inattention or arousal Hippocampal information process-ing is believed to be behind this system which isassociated with a higher degree of conscious aware-ness while the first two are mostly attributed tosubconscious processes

The cardiovascular system consists of the heart and allthe blood vessels and the variables of particular interestare BVP and HR BVP is the rate of flow of blood througha particular blood vessel and is related to both bloodpressure and the diameter of the vessel Constriction ordilation of the vessels is controlled by the sympatheticbranch of the ANS and along with electrodermal activityit has been shown to be related to information process-ing and decision making (Papillo amp Shapiro 1990) HRrefers to the frequency of the contractions of the heartmuscle or myocardium Specialized neuronsmdashthe so-called lsquolsquopacemakerrsquorsquo cellsmdashinitiate the contraction ofmyocardium and the output of the pacemakers iscontrolled by both parasympathetic (HR decrease) andsympathetic (HR increase) ANS branches BVP and HRtrack each other closely so that HR deceleration usuallycauses an increase in BVP Compared to SCR arousalindications that refer mostly to cognitive processeschanges in cardiovascular variables register higher levelsof arousal which are often somatically mediated Thesevariables may provide supplemental information forinterpreting high SCR signalsmdashelevated SCR accompa-nied by an increase in HR may be an indication ofextreme significance of the task or stimulus or alter-natively simply indicate that some physical activity isbeing performed simultaneously (Boucsein 1992)

EMG measurements are based on the electrical signalsgenerated by the contraction process of striatedmuscles Muscle action potentials travel along themuscle and some portion of the electrical activity leaksto the skin surface where it can be detected with thehelp of surface electrodes The activation of particularmuscles reflects different types of actions For exampleactivation of the facial muscle (ie the lsquolsquomasseterrsquorsquomuscle) indicates ongoing speech activation of theforearm muscle group (ie the lsquolsquoflexor digitorumrsquorsquogroup) corresponds to finger movements such as typingon a keyboard In addition to their role in motor activitycertain muscle groups exhibit strong correlations withemotional states Most of the research in this literaturehas focused on facial muscles since facial expressionshave been linked to emotional states (Ekman 1982) Forexample an increase in EMG activity over the forehead isassociated with anxiety and tension and an increase inactivity over the brow accompanied by a slight decreasein activity over the cheek often corresponds to unpleas-ant sensory stimuli (Cacioppo et al 2000) ThereforeEMG measurements can capture very subtle changes inmuscle activity that can differentiate otherwise indistin-guishable response patterns However in our currentimplementation the role of EMG measurements is

limited to identifying and eliminating anomalous sensorreadings caused by certain physical motions of a subject

Respiration influences the HR through vascular recep-tors (Lorig amp Schwartz 1990) and although this variableis usually not of primary psychological importance it is areliable indicator of physically demanding activitiesundertaken by the subject for example speaking orcoughing which can often yield anomalous sensor read-ings if not properly taken into account Respiration canbe measured by placing a sensor that monitors chestexpansion and compression

Finally body temperature regulation involves the in-tegration of autonomic motor and endocrine responsesand several studies have related the temperatures ofdifferent parts of the body to certain cognitive and emo-tional contents of the task or stimuli For example fore-head temperature (a proxy for brain temperature)increases while experiencing negative emotions coolingenhances positive affect while warming depresses it(McIntosh Zajonc Vig amp Emerick 1997) Another studyreports that hand skin temperature increases with pos-itive affect and decreases with threatening and unpleas-ant tasks (Rimm-Kaufman amp Kagan 1996)

RESULTS

Our sample of subjects consisted of 10 professionaltraders employed by the foreign-exchange and inter-est-rate derivatives business unit of a major globalfinancial institution based in Boston MA This institu-tion provides banking and other financial services toclients ranging from small regional startups to Fortune500 multinational corporations The foreign-exchangeand interest-rate derivatives business unit employs ap-proximately 90 professionals of which two-thirds spe-cializes in the trading of foreign exchange and relatedinstruments and one-third specializes in the trading ofinterest-rate derivative securities In a typical day thisbusiness unit engages in 1000- 1200 trades with anaverage size of US$3- 5 million per trade Approximately80 of the trades are executed on behalf of the clientsof the financial institution with the remaining 20motivated by the financial institutionrsquos market-makingactivities

For each of the 10 subjects five physiological variableswere monitored in real time during the entire duration ofeach session while the subjects sat at their trading con-soles (see Figure 1) Six sensors were used SCR BVPbody temperature (TMP) respiration (RSP) and twoelectromyographic sensors (facial and forearm EMG)

The benefits of acquiring real-time physiological meas-urements in vivo must be balanced against the cost ofmeasurement error spurious signals and other statisticalartifacts which if left untreated can obscure and con-found any genuine signals in the data To eliminate asmany artifacts as possible while maximizing the informa-tional content of the data each of the five physiological

Lo and Repin 325

variables were preprocessed with filters calibrated toeach variable individually and adaptive time-windowswere used in place of fixed-length windows to matchthe event-driven nature of the market variables

After preprocessing the following seven features wereextracted from SCR BVP temperature and respirationsignals

1 times of onset of SCR responses2 amplitudes of SCR responses3 average HR (every 3 sec)4 BVP signal amplitudes5 respiration rates (every 5 sec)6 respiration amplitudes7 temperature changes (from the 10-sec lag)

These features were selected to reflect the differentproperties and different information contained in theraw physiological variables SCRs were characterized bysmooth lsquolsquobumpsrsquorsquo with a relatively fast onset and slowdecay hence the number of such bumps and theirrelative strength are excellent summary measures ofthe SCR signal Intervals of 3 and 5 sec for heart andrespiration rates were chosen as a compromise betweenthe need for finer time slices to distinguish the onsetand completion of physiological and market events andthe necessity of a sufficient number of heartbeats andrespiration cycles within each interval to compute anaverage Under normal conditions a typical subject willexhibit an average of three to four heartbeats per 3-secinterval and approximately three breaths per 5-sec in-terval Temperature was the slowest-varying physiolog-ical signal and the 10-sec interval to register its changereflected the maximum possible interval in our analysis

For each session real-time market data for key finan-cial instruments actively traded or monitored by thesubject were collected synchronously with the physio-logical data In particular across the 10 subjects a total of15 instruments were consideredmdash13 foreign currencies

and two futures contracts the euro (EUR) the Japaneseyen (JPY) the British pound (GBP) the Canadian dollar(CAD) the Swiss franc (CHF) the Australian dollar(AUD) the New Zealand dollar (NZD) the Mexican peso(MXP) the Brazilian real (BRL) the Argentinian peso(ARS) the Colombian peso (COP) the Venezuelan boli-var (VEB) the Chilean peso (CLP) SampP 500 futures(SPU) and eurodollar futures (EDU) For each securityfour time series were monitored (1) the bid price Pt

b (2)the ask price Pt

a (3) the bidask spread Xt sup2 Ptb iexcl Pt

aand (4) the net return Rt sup2 (Pt iexcl Ptiexcl1) Ptiexcl1 where Pt

denotes the mid-spread price at time t and all priceswere sampled every second

For each of the financial time series we identifiedthree classes of market events deviations trend-rever-sals and volatility events These events are often cited bytraders as significant developments that require height-ened attention potentially signaling a shift in marketdynamics and risk exposures With these definitions andcalibrations in hand we implemented an automaticprocedure for detecting deviations trend-reversals andvolatility events for all the relevant time series in eachsession Table 1 reports event counts for the financialtime series monitored for each subject Feature vectorswere then constructed for all detected market events inall time series for all subjects and a set of control featurevectors was generated by applying the same feature-extraction process to randomly selected windows con-taining no events of any kind One set of control featurevectors was constructed for deviation and trend-reversalevents (10-sec intervals) and a second set was con-structed for volatility-type events (5-min intervals) Atwo-sided t test was applied to each component of eachfeature vector and the corresponding control vector totest the null hypothesis that the feature vectors werestatistically indistinguishable from the control featurevectors

For volatility events which by definition occurred in5-min intervals (event windows) we constructed featurevectors containing the following information for theduration of the event window

1 number of SCR responses2 average SCR amplitude3 average HR4 average ratio of the BVP amplitude to local baseline5 average ratio of the BVP amplitude to global

baseline6 number of temperature changes exceeding 018F7 average respiration rate8 average respiration amplitude

where the local baseline for the BVP signal is the averagelevel during the event window and the global baseline isthe average over the entire recording session for eachtrader Control feature vectors were constructed byapplying the feature-extraction process to randomlyselected 5-min windows containing no events of any

Laptop with data acquisition software

Fiber optic cable Control unit

Market information (Reuters Bloomberg) Order screen etc TIBCO market spreadsheet

eal-time data feed the market data

acquisition computer

Figure 1 Typical set-up for the measurement of real-time physiolo-gical responses of financial traders during live trading sessions Real-time market data used by the traders are recorded synchronously andsubsequently analyzed together with the physiological response data

326 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

kind For the other two types of market eventsmdashdevia-tions and trend-reversalsmdashlsquolsquopre-eventrsquorsquo and lsquolsquopost-eventrsquorsquofeature vectors were constructed before and after eachmarket event using the same variables where featureswere aggregated over 10-sec windows immediately pre-ceding and following the event Control feature vectorswere generated for these cases as well

The motivation for the post-event feature vector wasto capture the subjectrsquos reaction to the event with thepre-event feature vector as a benchmark from which tomeasure the magnitude of the reaction A comparisonbetween the pre-event feature vector and a controlfeature vector may provide an indication of a subjectrsquosanticipation of the event Latencies of the autonomicresponses reported in previous studies were on a timescale of 1 to 10 sec (see Cacioppo et al 2000) hence10-sec event windows were judged to be long enough forevent-related autonomic responses to occur and at thesame time short enough to minimize the likelihood ofoverlaps with other events or anomalies Five-minutewindows were used for volatility events because 10-secintervals were simply insufficient for meaningful volatilitycalculations For all of the financial time series used inthis study and for most financial time series in generalthere are few volatility events that occur in any 10-secinterval except of course under extreme conditions forexample the stock market crash of October 19 1987 (nosuch conditions prevailed during any of our sessions)

The statistical analysis of the physiological and marketdata was motivated by four objectives (1) to identifyparticular classes of events with statistically significantdifferences in mean autonomic responses before or after

an event as compared to the no-event control (2) toidentify particular traders or groups of traders based onexperience or other personal characteristics that dem-onstrate significant mean autonomic responses duringmarket events (3) to identify particular financial instru-ments or groups of instruments that are associated withsignificant mean autonomic responses and (4) to iden-tify particular physiological variables that demonstratesignificant mean response levels immediately followingthe event or during the event-anticipation period ascompared to the no-event control

To address the first objective a total of eight types ofevents were used for each of the time series

1 price deviations2 spread deviations3 return deviations4 price trend-reversals5 spread trend-reversals6 maximum volatility7 price volatility8 return volatility

and for each type of event a two-sided t test wasperformed for the pooled sample of all subjects and alltime series in which mean autonomic responses duringevent periods were compared to mean autonomic re-sponses during no-event control periods

The t statistics corresponding to the two-sided hypoth-esis test are given in the left subpanel of Table 2 labelledlsquolsquoAll Tradersrsquorsquo Statistics that are significant at the 5levelmdashbased on the t distribution with the appropriatedegrees of freedom for each entrymdashare highlighted For

Table 1 Number of Deviation (DEV) Trend-Reversal (TRV) and Volatility (VOL) Events Detected in Real-Time Market Data forEach Trader Over the Course of Each Trading Session

EUR JPYOther MajorCu rrenciesa

Latin Am ericanCu rrenciesb Derivativesc

Trader ID DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL

B32 21 16 15 25 17 15 - - - - - - - - -

B34 18 10 14 17 17 14 - - - - - - - - -

B35 20 16 14 13 17 15 21 9 15 24 16 13 - - -

B36 - - - - - - - - - - - - 28 23 31

B37 19 19 14 18 18 12 18 18 14 105 86 63 - - -

B38 21 20 12 22 18 13 23 20 15 96 61 56 - - -

B39 - - - - - - - - - - - - 40 37 23

B310 10 17 10 18 16 14 116 59 61 - - - - - -

B311 17 15 15 17 16 15 94 61 67 - - - - - -

B312 19 17 15 16 19 13 107 83 71 - - - - - -

aGBP CAD CHF AUD NZDbMXP BRL ARS COP VEB CLPcEDU SPU

Lo and Repin 327

Tab

le2

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vec

tors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

All

Tra

ders

Hig

hEx

peri

ence

Low

orM

oder

ate

Expe

rien

ce

Phys

iolo

gy

Fea

ture

s

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VOL

Ma

xim

um

DE

V

Pri

ce

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

Nu

mbe

ro

fSC

Rre

spon

ses

28

72

060

42

08

81

66

40

636

057

90

561

122

8iexcl

022

1iexcl

126

10

184

20

81

26

96

012

32

75

22

17

0iexcl

066

5iexcl

068

8

Aver

age

SCR

ampl

itud

eiexcl

069

50

887

018

6iexcl

247

81

136

088

00

030

034

40

744

iexcl1

340

114

90

904

075

61

578

iexcl1

265

iexcl1

947

iexcl0

012

iexcl0

615

Aver

age

HR

090

5iexcl

058

8iexcl

059

6iexcl

061

71

032

iexcl0

270

044

91

88

6iexcl

179

7iexcl

163

51

83

6iexcl

201

81

310

005

1iexcl

038

2iexcl

110

6iexcl

031

11

283

Aver

age

BV

Pam

plit

ude

rela

tive

tolo

cal

base

line

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9iexcl

006

1iexcl

074

70

334

001

50

164

26

23

iexcl0

204

iexcl1

775

087

51

79

6iexcl

128

92

70

8

Aver

age

BV

Pam

plit

ude

rela

tive

togl

obal

base

line

141

7iexcl

272

62

32

41

417

iexcl2

211

28

86

iexcl2

529

iexcl1

367

159

01

110

iexcl1

165

23

01

19

17

iexcl2

664

18

84

20

82

iexcl1

759

27

58

Nu

mbe

ro

fte

mp

erat

ure

jum

ps

098

82

82

00

650

17

77

48

68

iexcl2

482

iexcl1

010

iexcl1

435

NA

NA

NA

164

40

837

29

02

114

52

44

52

63

3iexcl

321

3

Aver

age

resp

irat

ion

rate

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4iexcl

155

61

109

012

10

866

iexcl0

102

009

30

079

iexcl0

552

032

3iexcl

063

1iexcl

217

0iexcl

152

8

Aver

age

resp

irat

ion

ampl

itud

e

iexcl1

646

iexcl0

250

013

60

034

iexcl1

487

iexcl3

499

108

70

265

iexcl1

777

iexcl0

657

012

1iexcl

060

7iexcl

274

9iexcl

131

92

14

20

055

iexcl2

832

iexcl4

533

Bo

lded

stat

isti

csar

esi

gnifi

can

tat

the

5le

vel

Th

ele

ftp

anel

give

sag

greg

ate

resu

lts

for

allt

rad

ers

Sim

ilar

ttes

tsfo

rtr

ader

sw

ith

hig

hex

per

ienc

ean

dlo

wo

rm

oder

ate

exp

erie

nce

are

show

nin

the

mid

dle

and

righ

tp

anel

sre

spec

tive

lyE

ntr

ies

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

wer

ep

rese

nt

inei

ther

the

even

tor

cont

rol

feat

ure

vect

or

for

the

par

ticu

lar

mar

ket-

even

tty

pe

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spo

nse

sH

R=

hea

rtra

te

BV

P=

blo

od

volu

me

pul

se

328 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

deviations and trend-reversals both the number of SCRsand the number of temperature jumps reached statisti-cally significant levels for three out of the five event typesBVP amplitude-related features were highly significantfor volatility eventsmdashsignificance levels less than 1 wereobtained for both BVP amplitude features for all threetypes of volatility events Because the results were sosimilar across the three types of volatility events (max-imum volatility price volatility and return volatility) wereport the results only for maximum volatility in Table 2Such patterns of autonomic responses may indicate thepresence of transient emotional responses to deviationsand trend-reversal events that occur within 10-sec win-dows while volatility events defined in 5-min windowselicited a tonic response in BVP

To explore potential differences between experi-enced and less experienced traders a sample of 10traders was divided into two groups each consistingof five traders the first containing highly experiencedtraders and the second containing traders with low ormoderate experience where the levels of experiencewere determined through interviews with the tradersrsquosupervisors The results for each of the two groupsare reported in the two right subpanels of Table 2labelled lsquolsquoHigh Experiencersquorsquo and lsquolsquoLow or ModerateExperiencersquorsquo The high-experience traders generallyexhibited low responses for deviations and trend-re-versal events with only two types of market events(spread deviations and spread trend-reversals) elicitingsignificant mean autonomic responses (average HR)However even for high-experience traders volatilityevents induced statistically significant mean responsesfor both SCR and BVP amplitude Less experiencedtraders showed a much higher number of significantmean responses in the number of SCRs BVP ampli-tude and number of temperature increases for devia-tions and trend-reversal events Table 2 shows thatsessions with low- and moderate-experience tradersyielded 13 t statistics significant at the 5 level whilesessions with high-experience traders yielded only fivet statistics significant at the 5 level For both sets oftraders volatility events were associated with signifi-cant BVP amplitude The differences in responsepatterns for deviations and trend-reversals observedin this case indicate that less experienced traders maybe more sensitive to short-term changes in the marketvariables than their more experienced colleagues

To address the third objective 10-sec intervals imme-diately preceding and immediately following each devia-tion and trend-reversal event were compared to control(ie no-event) time intervals Two separate t tests wereconductedmdashone for pre- and another for post-eventtime intervalsmdashfor each of the three types of deviationsand two types of trend-reversals Because the objectivewas to detect differences in how pre- and post-eventfeature vectors differed from the control we used thesame control feature vectors for both sets of t tests In all

cases the t tests were designed to test the null hypoth-esis that both pre- and post-event feature vectors werestatistically indistinguishable from the control featurevector Surprisingly these two sets of t tests yielded verysimilar results reported in Table 3 implying that noneof the physiological variables were predictors of antici-patory emotional responses These findings may be atleast partly explained by how the events were defined Inparticular the definitions of deviation and trend-reversalevents are those instances where the time seriesachieved a prespecified deviation from the time-seriesmean and its moving average respectively In theabsence of large jumps the values of the time seriesnear an event defined in this way are likely to becomparable to the event itself Therefore the tradersmay be responding to market conditions occurringthroughout the pre- and post-event period not just atthe exact time of the event hence the similarity be-tween the two periods More complex definitions ofevents may allow us to discriminate between pre- andpost-event physiological responses and we are explor-ing several alternatives in ongoing research

Finally to address the fourth objective the followingfour groups of financial instruments were formed on thebasis of similarity in their statistical characteristics

Group 1 EUR JPY

Group 2 GBP CAD CHF AUD NZD (other majorcurrencies)

Group 3 MXP BRL ARS COP VEB CLP (LatinAmerican currencies)

Group 4 SPU EDU (derivatives)

Group 1 is by far the most actively traded set offinancial instruments and accounts for most of thetrading activities of our sample of 10 traders hence itis not surprising that the pattern of statistical signifi-cance for Group 1 in Table 4 is almost identical to thepattern for all traders in Table 2 (the left panel) SCRis significant for price and return deviations temper-ature jumps are significant for spread and price devia-tions and both measures of BVP are significant forvolatility events For Group 2 none of the eightphysiological variables were significant for price orspread deviations although temperature changes andrespiration rates were significant for return deviationsand both BVP measures were significant for pricetrend-reversals and volatility events The financial in-struments in Group 2 were not the primary focus ofany of the traders in our sample which may explainthe different patterns of significance as compared withGroup 1 Group 4 contains the fewest significantresponses and this is consistent with the fact thatthese securities were traded by two of the mostexperienced traders in our sample Derivative secur-ities are considerably more complex instruments thanthe spot currencies requiring more skill and trainingthan the other groups of securities However SCR was

Lo and Repin 329

Tab

le3

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gPr

e-an

dPo

st-e

vent

Tim

eIn

terv

als

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

Pre

-eve

nt

Inte

rva

lP

ost-

even

tIn

terv

al

Phy

sio

logy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mbe

rof

SCR

resp

ons

es3

75

61

543

25

40

21

18

24

01

057

92

87

20

604

20

88

16

64

063

60

579

Aver

age

SCR

amp

litu

de0

700

iexcl0

454

iexcl1

793

051

11

068

088

0iexcl

069

50

887

018

6iexcl

247

81

136

088

0

Aver

age

HR

096

8iexcl

020

5iexcl

055

2iexcl

066

3iexcl

038

6iexcl

027

00

905

iexcl0

588

iexcl0

596

iexcl0

617

103

2iexcl

027

0

Aver

age

BVP

amp

litud

ere

lativ

eto

loca

lba

selin

e0

043

iexcl0

895

006

00

814

iexcl0

045

36

19

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9

Aver

age

BVP

amp

litud

ere

lativ

eto

glob

alba

selin

e1

183

iexcl2

794

23

97

131

7iexcl

225

62

88

61

417

iexcl2

726

23

24

141

7iexcl

221

12

88

6

Nu

mbe

rof

tem

per

atur

eju

mp

s1

110

27

67

050

52

15

83

92

5iexcl

248

20

988

28

20

065

01

77

74

86

8iexcl

248

2

Aver

age

resp

irat

ion

rate

126

1iexcl

165

0iexcl

025

41

303

iexcl1

999

iexcl0

064

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4

Aver

age

resp

irat

ion

amp

litud

eiexcl

162

0iexcl

006

70

561

006

8iexcl

180

7iexcl

349

9iexcl

164

6iexcl

025

00

136

003

4iexcl

148

7iexcl

349

9

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

l

Th

ele

ftp

anel

cont

ain

st

stat

isti

csfo

rp

re-e

vent

feat

ure

vect

ors

and

the

righ

tp

anel

con

tain

st

stat

isti

csfo

rp

ost-

even

tfe

atu

reve

cto

rs

both

test

edag

ain

stth

esa

me

set

of

con

tro

ls

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spon

ses

HR

=h

eart

rate

B

VP

=bl

ood

volu

me

pu

lse

330 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

financial risk measures and emotional states and dynam-ics In a pilot sample of 10 traders we find statisticallysignificant differences in mean electrodermal responsesduring transient market events as compared with no-event control baselines and statistically significant meanchanges in cardiovascular variables during periods ofheightened market volatility as compared with normal-volatility control baselines We also observe significantdifferences in mean physiological responses among the10 traders that may be systematically related to theamount of trading experience

In studying the link between emotion and rationaldecision making in the face of uncertainty financialsecurities traders are ideal subjects for several reasonsBecause the basic functions of securities trading involvefrequent decisions concerning riskreward trade-offstraders are almost continuously engaged in the activitythat we wish to study This allows us to conduct ourstudy in vivo and with minimal interference to andtherefore minimal contamination of the subjectsrsquo natu-ral motives and behavior Traders are typically providedwith significant economic incentives to avoid many ofthe biases that are often associated with irrational invest-ment practices Moreover they are highly paid profes-sionals that have undergone a variety of trainingexercises apprenticeships and trial periods throughwhich their skills have been finely honed Thereforethey are likely to be among the most rational decisionmakers in the general population hence ideal subjectsfor examining the role of emotion in rational decision-making processes Finally due to the real-time nature ofmost professional trading operations it is possible toconstruct accurate real-time records of the environmentin which traders make their decisions namely thefluctuations of market prices of the securities that theytrade With such real-time records we are able to matchmarket events such as periods of high price-volatilitywith synchronous real-time measurements of physiolog-ical characteristics

To measure the emotional responses of our subjectsduring their trading activities we focus on indirectmanifestations through the responses of the autonomicnervous system (ANS) (Cacioppo Tassinary amp Bernt2000) The ANS innervates the viscera and is responsiblefor the regulation of internal states that are mediated byinternal bodily as well as emotional and cognitiveprocesses ANS responses are relatively easy to measuresince many of them can be measured noninvasively fromexternal body sites without interfering with cognitivetasks performed by the subject ANS responses occur onthe scale of seconds which is essential for investigationof real-time risk processing

Previous studies have focused primarily on time-aver-aged (over hours or tens of minutes) levels of autonomicactivity as a function of task complexity or mental strainFor example some experiments have considered thelink between autonomic activity and driving conditions

and road familiarity for nonprofessional drivers (Brownamp Huffman 1972) and the stage of flight (take-offsteady flight landing) in a jetfighter flight simulator(Lindholm amp Cheatham 1983) Recently the focus ofresearch has started to shift towards finer temporalscales (seconds) of autonomic responses associated withcognitive and emotional processes Perhaps the mostinfluential set of experiments in this area was conductedin the broad context of an investigation of the role ofemotion in decision-making processes (Damasio 1994)In one of these experiments skin conductance re-sponses (SCRs) were measured in subjects involved ina gambling task (Bechara Damasio Tranel amp Damasio1997) The results indicated that the anticipation of themore risky outcomes led to more SCRs than of the lessrisky ones The brain circuitry involved in anticipatingmonetary rewards has also been localized (BreiterAharon Kahneman Dale amp Shizgal 2001) Anotherstudy (Frederikson et al 1998) reported neuroanatom-ical correlates of skin conductance activity with the brainregions that also support anticipation affect and cogni-tively or emotionally mediated motor preparation Re-cent experiments (Critchley Krase amp Statman 1999)support the significance of autonomic responses duringrisk-taking and reward-related behavior They providemore details on brain activation correlates of peripheralautonomic responses and also claim the possibility ofdiscriminating the activity patterns related to changingversus continuing the behavior based on the immediategainloss history in a gambling task

We focus on five types of physiological characteristicsin our study skin conductance cardiovascular data (BVPand HR) electromyographic (EMG) data respirationrate and body temperature

SCR is measured by the voltage drop between twoelectrodes placed on the skin surface a few centimetersapart Changes in skin conductance occur when eccrinesweat glands that are innervated by the sympathetic ANSfibers receive a signal from a certain part of the brainThree distinct lsquolsquobrain information systemsrsquorsquo can poten-tially elicit SCR signals (Boucsein 1992) The affectarousal system is capable of generating a phasic re-sponse (on the scale of seconds) that is associateddirectly with the sensory input indicating attentionfocusing or defense response and the amygdala is theprimary brain region involved The emotional responseto a novel or highly complex task is an example of affectarousal Another system that can initiate a phasic SCRcenters on the basal ganglia and is related to a prepar-atory activation mediated by internal cognitive pro-cesses The body exhibits increased perceptual andmotor readiness and high attention levels Expectationof an event or preparation to an important actionillustrates the operation of this system The third systemoften called the lsquolsquoeffort systemrsquorsquo produces tonic changesin the level of skin conductance and is related to thelong-term changes of a general emotional (hedonic)

324 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

state or attitude indicating a nonspecific increase inattention or arousal Hippocampal information process-ing is believed to be behind this system which isassociated with a higher degree of conscious aware-ness while the first two are mostly attributed tosubconscious processes

The cardiovascular system consists of the heart and allthe blood vessels and the variables of particular interestare BVP and HR BVP is the rate of flow of blood througha particular blood vessel and is related to both bloodpressure and the diameter of the vessel Constriction ordilation of the vessels is controlled by the sympatheticbranch of the ANS and along with electrodermal activityit has been shown to be related to information process-ing and decision making (Papillo amp Shapiro 1990) HRrefers to the frequency of the contractions of the heartmuscle or myocardium Specialized neuronsmdashthe so-called lsquolsquopacemakerrsquorsquo cellsmdashinitiate the contraction ofmyocardium and the output of the pacemakers iscontrolled by both parasympathetic (HR decrease) andsympathetic (HR increase) ANS branches BVP and HRtrack each other closely so that HR deceleration usuallycauses an increase in BVP Compared to SCR arousalindications that refer mostly to cognitive processeschanges in cardiovascular variables register higher levelsof arousal which are often somatically mediated Thesevariables may provide supplemental information forinterpreting high SCR signalsmdashelevated SCR accompa-nied by an increase in HR may be an indication ofextreme significance of the task or stimulus or alter-natively simply indicate that some physical activity isbeing performed simultaneously (Boucsein 1992)

EMG measurements are based on the electrical signalsgenerated by the contraction process of striatedmuscles Muscle action potentials travel along themuscle and some portion of the electrical activity leaksto the skin surface where it can be detected with thehelp of surface electrodes The activation of particularmuscles reflects different types of actions For exampleactivation of the facial muscle (ie the lsquolsquomasseterrsquorsquomuscle) indicates ongoing speech activation of theforearm muscle group (ie the lsquolsquoflexor digitorumrsquorsquogroup) corresponds to finger movements such as typingon a keyboard In addition to their role in motor activitycertain muscle groups exhibit strong correlations withemotional states Most of the research in this literaturehas focused on facial muscles since facial expressionshave been linked to emotional states (Ekman 1982) Forexample an increase in EMG activity over the forehead isassociated with anxiety and tension and an increase inactivity over the brow accompanied by a slight decreasein activity over the cheek often corresponds to unpleas-ant sensory stimuli (Cacioppo et al 2000) ThereforeEMG measurements can capture very subtle changes inmuscle activity that can differentiate otherwise indistin-guishable response patterns However in our currentimplementation the role of EMG measurements is

limited to identifying and eliminating anomalous sensorreadings caused by certain physical motions of a subject

Respiration influences the HR through vascular recep-tors (Lorig amp Schwartz 1990) and although this variableis usually not of primary psychological importance it is areliable indicator of physically demanding activitiesundertaken by the subject for example speaking orcoughing which can often yield anomalous sensor read-ings if not properly taken into account Respiration canbe measured by placing a sensor that monitors chestexpansion and compression

Finally body temperature regulation involves the in-tegration of autonomic motor and endocrine responsesand several studies have related the temperatures ofdifferent parts of the body to certain cognitive and emo-tional contents of the task or stimuli For example fore-head temperature (a proxy for brain temperature)increases while experiencing negative emotions coolingenhances positive affect while warming depresses it(McIntosh Zajonc Vig amp Emerick 1997) Another studyreports that hand skin temperature increases with pos-itive affect and decreases with threatening and unpleas-ant tasks (Rimm-Kaufman amp Kagan 1996)

RESULTS

Our sample of subjects consisted of 10 professionaltraders employed by the foreign-exchange and inter-est-rate derivatives business unit of a major globalfinancial institution based in Boston MA This institu-tion provides banking and other financial services toclients ranging from small regional startups to Fortune500 multinational corporations The foreign-exchangeand interest-rate derivatives business unit employs ap-proximately 90 professionals of which two-thirds spe-cializes in the trading of foreign exchange and relatedinstruments and one-third specializes in the trading ofinterest-rate derivative securities In a typical day thisbusiness unit engages in 1000- 1200 trades with anaverage size of US$3- 5 million per trade Approximately80 of the trades are executed on behalf of the clientsof the financial institution with the remaining 20motivated by the financial institutionrsquos market-makingactivities

For each of the 10 subjects five physiological variableswere monitored in real time during the entire duration ofeach session while the subjects sat at their trading con-soles (see Figure 1) Six sensors were used SCR BVPbody temperature (TMP) respiration (RSP) and twoelectromyographic sensors (facial and forearm EMG)

The benefits of acquiring real-time physiological meas-urements in vivo must be balanced against the cost ofmeasurement error spurious signals and other statisticalartifacts which if left untreated can obscure and con-found any genuine signals in the data To eliminate asmany artifacts as possible while maximizing the informa-tional content of the data each of the five physiological

Lo and Repin 325

variables were preprocessed with filters calibrated toeach variable individually and adaptive time-windowswere used in place of fixed-length windows to matchthe event-driven nature of the market variables

After preprocessing the following seven features wereextracted from SCR BVP temperature and respirationsignals

1 times of onset of SCR responses2 amplitudes of SCR responses3 average HR (every 3 sec)4 BVP signal amplitudes5 respiration rates (every 5 sec)6 respiration amplitudes7 temperature changes (from the 10-sec lag)

These features were selected to reflect the differentproperties and different information contained in theraw physiological variables SCRs were characterized bysmooth lsquolsquobumpsrsquorsquo with a relatively fast onset and slowdecay hence the number of such bumps and theirrelative strength are excellent summary measures ofthe SCR signal Intervals of 3 and 5 sec for heart andrespiration rates were chosen as a compromise betweenthe need for finer time slices to distinguish the onsetand completion of physiological and market events andthe necessity of a sufficient number of heartbeats andrespiration cycles within each interval to compute anaverage Under normal conditions a typical subject willexhibit an average of three to four heartbeats per 3-secinterval and approximately three breaths per 5-sec in-terval Temperature was the slowest-varying physiolog-ical signal and the 10-sec interval to register its changereflected the maximum possible interval in our analysis

For each session real-time market data for key finan-cial instruments actively traded or monitored by thesubject were collected synchronously with the physio-logical data In particular across the 10 subjects a total of15 instruments were consideredmdash13 foreign currencies

and two futures contracts the euro (EUR) the Japaneseyen (JPY) the British pound (GBP) the Canadian dollar(CAD) the Swiss franc (CHF) the Australian dollar(AUD) the New Zealand dollar (NZD) the Mexican peso(MXP) the Brazilian real (BRL) the Argentinian peso(ARS) the Colombian peso (COP) the Venezuelan boli-var (VEB) the Chilean peso (CLP) SampP 500 futures(SPU) and eurodollar futures (EDU) For each securityfour time series were monitored (1) the bid price Pt

b (2)the ask price Pt

a (3) the bidask spread Xt sup2 Ptb iexcl Pt

aand (4) the net return Rt sup2 (Pt iexcl Ptiexcl1) Ptiexcl1 where Pt

denotes the mid-spread price at time t and all priceswere sampled every second

For each of the financial time series we identifiedthree classes of market events deviations trend-rever-sals and volatility events These events are often cited bytraders as significant developments that require height-ened attention potentially signaling a shift in marketdynamics and risk exposures With these definitions andcalibrations in hand we implemented an automaticprocedure for detecting deviations trend-reversals andvolatility events for all the relevant time series in eachsession Table 1 reports event counts for the financialtime series monitored for each subject Feature vectorswere then constructed for all detected market events inall time series for all subjects and a set of control featurevectors was generated by applying the same feature-extraction process to randomly selected windows con-taining no events of any kind One set of control featurevectors was constructed for deviation and trend-reversalevents (10-sec intervals) and a second set was con-structed for volatility-type events (5-min intervals) Atwo-sided t test was applied to each component of eachfeature vector and the corresponding control vector totest the null hypothesis that the feature vectors werestatistically indistinguishable from the control featurevectors

For volatility events which by definition occurred in5-min intervals (event windows) we constructed featurevectors containing the following information for theduration of the event window

1 number of SCR responses2 average SCR amplitude3 average HR4 average ratio of the BVP amplitude to local baseline5 average ratio of the BVP amplitude to global

baseline6 number of temperature changes exceeding 018F7 average respiration rate8 average respiration amplitude

where the local baseline for the BVP signal is the averagelevel during the event window and the global baseline isthe average over the entire recording session for eachtrader Control feature vectors were constructed byapplying the feature-extraction process to randomlyselected 5-min windows containing no events of any

Laptop with data acquisition software

Fiber optic cable Control unit

Market information (Reuters Bloomberg) Order screen etc TIBCO market spreadsheet

eal-time data feed the market data

acquisition computer

Figure 1 Typical set-up for the measurement of real-time physiolo-gical responses of financial traders during live trading sessions Real-time market data used by the traders are recorded synchronously andsubsequently analyzed together with the physiological response data

326 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

kind For the other two types of market eventsmdashdevia-tions and trend-reversalsmdashlsquolsquopre-eventrsquorsquo and lsquolsquopost-eventrsquorsquofeature vectors were constructed before and after eachmarket event using the same variables where featureswere aggregated over 10-sec windows immediately pre-ceding and following the event Control feature vectorswere generated for these cases as well

The motivation for the post-event feature vector wasto capture the subjectrsquos reaction to the event with thepre-event feature vector as a benchmark from which tomeasure the magnitude of the reaction A comparisonbetween the pre-event feature vector and a controlfeature vector may provide an indication of a subjectrsquosanticipation of the event Latencies of the autonomicresponses reported in previous studies were on a timescale of 1 to 10 sec (see Cacioppo et al 2000) hence10-sec event windows were judged to be long enough forevent-related autonomic responses to occur and at thesame time short enough to minimize the likelihood ofoverlaps with other events or anomalies Five-minutewindows were used for volatility events because 10-secintervals were simply insufficient for meaningful volatilitycalculations For all of the financial time series used inthis study and for most financial time series in generalthere are few volatility events that occur in any 10-secinterval except of course under extreme conditions forexample the stock market crash of October 19 1987 (nosuch conditions prevailed during any of our sessions)

The statistical analysis of the physiological and marketdata was motivated by four objectives (1) to identifyparticular classes of events with statistically significantdifferences in mean autonomic responses before or after

an event as compared to the no-event control (2) toidentify particular traders or groups of traders based onexperience or other personal characteristics that dem-onstrate significant mean autonomic responses duringmarket events (3) to identify particular financial instru-ments or groups of instruments that are associated withsignificant mean autonomic responses and (4) to iden-tify particular physiological variables that demonstratesignificant mean response levels immediately followingthe event or during the event-anticipation period ascompared to the no-event control

To address the first objective a total of eight types ofevents were used for each of the time series

1 price deviations2 spread deviations3 return deviations4 price trend-reversals5 spread trend-reversals6 maximum volatility7 price volatility8 return volatility

and for each type of event a two-sided t test wasperformed for the pooled sample of all subjects and alltime series in which mean autonomic responses duringevent periods were compared to mean autonomic re-sponses during no-event control periods

The t statistics corresponding to the two-sided hypoth-esis test are given in the left subpanel of Table 2 labelledlsquolsquoAll Tradersrsquorsquo Statistics that are significant at the 5levelmdashbased on the t distribution with the appropriatedegrees of freedom for each entrymdashare highlighted For

Table 1 Number of Deviation (DEV) Trend-Reversal (TRV) and Volatility (VOL) Events Detected in Real-Time Market Data forEach Trader Over the Course of Each Trading Session

EUR JPYOther MajorCu rrenciesa

Latin Am ericanCu rrenciesb Derivativesc

Trader ID DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL

B32 21 16 15 25 17 15 - - - - - - - - -

B34 18 10 14 17 17 14 - - - - - - - - -

B35 20 16 14 13 17 15 21 9 15 24 16 13 - - -

B36 - - - - - - - - - - - - 28 23 31

B37 19 19 14 18 18 12 18 18 14 105 86 63 - - -

B38 21 20 12 22 18 13 23 20 15 96 61 56 - - -

B39 - - - - - - - - - - - - 40 37 23

B310 10 17 10 18 16 14 116 59 61 - - - - - -

B311 17 15 15 17 16 15 94 61 67 - - - - - -

B312 19 17 15 16 19 13 107 83 71 - - - - - -

aGBP CAD CHF AUD NZDbMXP BRL ARS COP VEB CLPcEDU SPU

Lo and Repin 327

Tab

le2

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vec

tors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

All

Tra

ders

Hig

hEx

peri

ence

Low

orM

oder

ate

Expe

rien

ce

Phys

iolo

gy

Fea

ture

s

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VOL

Ma

xim

um

DE

V

Pri

ce

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

Nu

mbe

ro

fSC

Rre

spon

ses

28

72

060

42

08

81

66

40

636

057

90

561

122

8iexcl

022

1iexcl

126

10

184

20

81

26

96

012

32

75

22

17

0iexcl

066

5iexcl

068

8

Aver

age

SCR

ampl

itud

eiexcl

069

50

887

018

6iexcl

247

81

136

088

00

030

034

40

744

iexcl1

340

114

90

904

075

61

578

iexcl1

265

iexcl1

947

iexcl0

012

iexcl0

615

Aver

age

HR

090

5iexcl

058

8iexcl

059

6iexcl

061

71

032

iexcl0

270

044

91

88

6iexcl

179

7iexcl

163

51

83

6iexcl

201

81

310

005

1iexcl

038

2iexcl

110

6iexcl

031

11

283

Aver

age

BV

Pam

plit

ude

rela

tive

tolo

cal

base

line

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9iexcl

006

1iexcl

074

70

334

001

50

164

26

23

iexcl0

204

iexcl1

775

087

51

79

6iexcl

128

92

70

8

Aver

age

BV

Pam

plit

ude

rela

tive

togl

obal

base

line

141

7iexcl

272

62

32

41

417

iexcl2

211

28

86

iexcl2

529

iexcl1

367

159

01

110

iexcl1

165

23

01

19

17

iexcl2

664

18

84

20

82

iexcl1

759

27

58

Nu

mbe

ro

fte

mp

erat

ure

jum

ps

098

82

82

00

650

17

77

48

68

iexcl2

482

iexcl1

010

iexcl1

435

NA

NA

NA

164

40

837

29

02

114

52

44

52

63

3iexcl

321

3

Aver

age

resp

irat

ion

rate

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4iexcl

155

61

109

012

10

866

iexcl0

102

009

30

079

iexcl0

552

032

3iexcl

063

1iexcl

217

0iexcl

152

8

Aver

age

resp

irat

ion

ampl

itud

e

iexcl1

646

iexcl0

250

013

60

034

iexcl1

487

iexcl3

499

108

70

265

iexcl1

777

iexcl0

657

012

1iexcl

060

7iexcl

274

9iexcl

131

92

14

20

055

iexcl2

832

iexcl4

533

Bo

lded

stat

isti

csar

esi

gnifi

can

tat

the

5le

vel

Th

ele

ftp

anel

give

sag

greg

ate

resu

lts

for

allt

rad

ers

Sim

ilar

ttes

tsfo

rtr

ader

sw

ith

hig

hex

per

ienc

ean

dlo

wo

rm

oder

ate

exp

erie

nce

are

show

nin

the

mid

dle

and

righ

tp

anel

sre

spec

tive

lyE

ntr

ies

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

wer

ep

rese

nt

inei

ther

the

even

tor

cont

rol

feat

ure

vect

or

for

the

par

ticu

lar

mar

ket-

even

tty

pe

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spo

nse

sH

R=

hea

rtra

te

BV

P=

blo

od

volu

me

pul

se

328 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

deviations and trend-reversals both the number of SCRsand the number of temperature jumps reached statisti-cally significant levels for three out of the five event typesBVP amplitude-related features were highly significantfor volatility eventsmdashsignificance levels less than 1 wereobtained for both BVP amplitude features for all threetypes of volatility events Because the results were sosimilar across the three types of volatility events (max-imum volatility price volatility and return volatility) wereport the results only for maximum volatility in Table 2Such patterns of autonomic responses may indicate thepresence of transient emotional responses to deviationsand trend-reversal events that occur within 10-sec win-dows while volatility events defined in 5-min windowselicited a tonic response in BVP

To explore potential differences between experi-enced and less experienced traders a sample of 10traders was divided into two groups each consistingof five traders the first containing highly experiencedtraders and the second containing traders with low ormoderate experience where the levels of experiencewere determined through interviews with the tradersrsquosupervisors The results for each of the two groupsare reported in the two right subpanels of Table 2labelled lsquolsquoHigh Experiencersquorsquo and lsquolsquoLow or ModerateExperiencersquorsquo The high-experience traders generallyexhibited low responses for deviations and trend-re-versal events with only two types of market events(spread deviations and spread trend-reversals) elicitingsignificant mean autonomic responses (average HR)However even for high-experience traders volatilityevents induced statistically significant mean responsesfor both SCR and BVP amplitude Less experiencedtraders showed a much higher number of significantmean responses in the number of SCRs BVP ampli-tude and number of temperature increases for devia-tions and trend-reversal events Table 2 shows thatsessions with low- and moderate-experience tradersyielded 13 t statistics significant at the 5 level whilesessions with high-experience traders yielded only fivet statistics significant at the 5 level For both sets oftraders volatility events were associated with signifi-cant BVP amplitude The differences in responsepatterns for deviations and trend-reversals observedin this case indicate that less experienced traders maybe more sensitive to short-term changes in the marketvariables than their more experienced colleagues

To address the third objective 10-sec intervals imme-diately preceding and immediately following each devia-tion and trend-reversal event were compared to control(ie no-event) time intervals Two separate t tests wereconductedmdashone for pre- and another for post-eventtime intervalsmdashfor each of the three types of deviationsand two types of trend-reversals Because the objectivewas to detect differences in how pre- and post-eventfeature vectors differed from the control we used thesame control feature vectors for both sets of t tests In all

cases the t tests were designed to test the null hypoth-esis that both pre- and post-event feature vectors werestatistically indistinguishable from the control featurevector Surprisingly these two sets of t tests yielded verysimilar results reported in Table 3 implying that noneof the physiological variables were predictors of antici-patory emotional responses These findings may be atleast partly explained by how the events were defined Inparticular the definitions of deviation and trend-reversalevents are those instances where the time seriesachieved a prespecified deviation from the time-seriesmean and its moving average respectively In theabsence of large jumps the values of the time seriesnear an event defined in this way are likely to becomparable to the event itself Therefore the tradersmay be responding to market conditions occurringthroughout the pre- and post-event period not just atthe exact time of the event hence the similarity be-tween the two periods More complex definitions ofevents may allow us to discriminate between pre- andpost-event physiological responses and we are explor-ing several alternatives in ongoing research

Finally to address the fourth objective the followingfour groups of financial instruments were formed on thebasis of similarity in their statistical characteristics

Group 1 EUR JPY

Group 2 GBP CAD CHF AUD NZD (other majorcurrencies)

Group 3 MXP BRL ARS COP VEB CLP (LatinAmerican currencies)

Group 4 SPU EDU (derivatives)

Group 1 is by far the most actively traded set offinancial instruments and accounts for most of thetrading activities of our sample of 10 traders hence itis not surprising that the pattern of statistical signifi-cance for Group 1 in Table 4 is almost identical to thepattern for all traders in Table 2 (the left panel) SCRis significant for price and return deviations temper-ature jumps are significant for spread and price devia-tions and both measures of BVP are significant forvolatility events For Group 2 none of the eightphysiological variables were significant for price orspread deviations although temperature changes andrespiration rates were significant for return deviationsand both BVP measures were significant for pricetrend-reversals and volatility events The financial in-struments in Group 2 were not the primary focus ofany of the traders in our sample which may explainthe different patterns of significance as compared withGroup 1 Group 4 contains the fewest significantresponses and this is consistent with the fact thatthese securities were traded by two of the mostexperienced traders in our sample Derivative secur-ities are considerably more complex instruments thanthe spot currencies requiring more skill and trainingthan the other groups of securities However SCR was

Lo and Repin 329

Tab

le3

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gPr

e-an

dPo

st-e

vent

Tim

eIn

terv

als

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

Pre

-eve

nt

Inte

rva

lP

ost-

even

tIn

terv

al

Phy

sio

logy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mbe

rof

SCR

resp

ons

es3

75

61

543

25

40

21

18

24

01

057

92

87

20

604

20

88

16

64

063

60

579

Aver

age

SCR

amp

litu

de0

700

iexcl0

454

iexcl1

793

051

11

068

088

0iexcl

069

50

887

018

6iexcl

247

81

136

088

0

Aver

age

HR

096

8iexcl

020

5iexcl

055

2iexcl

066

3iexcl

038

6iexcl

027

00

905

iexcl0

588

iexcl0

596

iexcl0

617

103

2iexcl

027

0

Aver

age

BVP

amp

litud

ere

lativ

eto

loca

lba

selin

e0

043

iexcl0

895

006

00

814

iexcl0

045

36

19

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9

Aver

age

BVP

amp

litud

ere

lativ

eto

glob

alba

selin

e1

183

iexcl2

794

23

97

131

7iexcl

225

62

88

61

417

iexcl2

726

23

24

141

7iexcl

221

12

88

6

Nu

mbe

rof

tem

per

atur

eju

mp

s1

110

27

67

050

52

15

83

92

5iexcl

248

20

988

28

20

065

01

77

74

86

8iexcl

248

2

Aver

age

resp

irat

ion

rate

126

1iexcl

165

0iexcl

025

41

303

iexcl1

999

iexcl0

064

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4

Aver

age

resp

irat

ion

amp

litud

eiexcl

162

0iexcl

006

70

561

006

8iexcl

180

7iexcl

349

9iexcl

164

6iexcl

025

00

136

003

4iexcl

148

7iexcl

349

9

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

l

Th

ele

ftp

anel

cont

ain

st

stat

isti

csfo

rp

re-e

vent

feat

ure

vect

ors

and

the

righ

tp

anel

con

tain

st

stat

isti

csfo

rp

ost-

even

tfe

atu

reve

cto

rs

both

test

edag

ain

stth

esa

me

set

of

con

tro

ls

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spon

ses

HR

=h

eart

rate

B

VP

=bl

ood

volu

me

pu

lse

330 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

state or attitude indicating a nonspecific increase inattention or arousal Hippocampal information process-ing is believed to be behind this system which isassociated with a higher degree of conscious aware-ness while the first two are mostly attributed tosubconscious processes

The cardiovascular system consists of the heart and allthe blood vessels and the variables of particular interestare BVP and HR BVP is the rate of flow of blood througha particular blood vessel and is related to both bloodpressure and the diameter of the vessel Constriction ordilation of the vessels is controlled by the sympatheticbranch of the ANS and along with electrodermal activityit has been shown to be related to information process-ing and decision making (Papillo amp Shapiro 1990) HRrefers to the frequency of the contractions of the heartmuscle or myocardium Specialized neuronsmdashthe so-called lsquolsquopacemakerrsquorsquo cellsmdashinitiate the contraction ofmyocardium and the output of the pacemakers iscontrolled by both parasympathetic (HR decrease) andsympathetic (HR increase) ANS branches BVP and HRtrack each other closely so that HR deceleration usuallycauses an increase in BVP Compared to SCR arousalindications that refer mostly to cognitive processeschanges in cardiovascular variables register higher levelsof arousal which are often somatically mediated Thesevariables may provide supplemental information forinterpreting high SCR signalsmdashelevated SCR accompa-nied by an increase in HR may be an indication ofextreme significance of the task or stimulus or alter-natively simply indicate that some physical activity isbeing performed simultaneously (Boucsein 1992)

EMG measurements are based on the electrical signalsgenerated by the contraction process of striatedmuscles Muscle action potentials travel along themuscle and some portion of the electrical activity leaksto the skin surface where it can be detected with thehelp of surface electrodes The activation of particularmuscles reflects different types of actions For exampleactivation of the facial muscle (ie the lsquolsquomasseterrsquorsquomuscle) indicates ongoing speech activation of theforearm muscle group (ie the lsquolsquoflexor digitorumrsquorsquogroup) corresponds to finger movements such as typingon a keyboard In addition to their role in motor activitycertain muscle groups exhibit strong correlations withemotional states Most of the research in this literaturehas focused on facial muscles since facial expressionshave been linked to emotional states (Ekman 1982) Forexample an increase in EMG activity over the forehead isassociated with anxiety and tension and an increase inactivity over the brow accompanied by a slight decreasein activity over the cheek often corresponds to unpleas-ant sensory stimuli (Cacioppo et al 2000) ThereforeEMG measurements can capture very subtle changes inmuscle activity that can differentiate otherwise indistin-guishable response patterns However in our currentimplementation the role of EMG measurements is

limited to identifying and eliminating anomalous sensorreadings caused by certain physical motions of a subject

Respiration influences the HR through vascular recep-tors (Lorig amp Schwartz 1990) and although this variableis usually not of primary psychological importance it is areliable indicator of physically demanding activitiesundertaken by the subject for example speaking orcoughing which can often yield anomalous sensor read-ings if not properly taken into account Respiration canbe measured by placing a sensor that monitors chestexpansion and compression

Finally body temperature regulation involves the in-tegration of autonomic motor and endocrine responsesand several studies have related the temperatures ofdifferent parts of the body to certain cognitive and emo-tional contents of the task or stimuli For example fore-head temperature (a proxy for brain temperature)increases while experiencing negative emotions coolingenhances positive affect while warming depresses it(McIntosh Zajonc Vig amp Emerick 1997) Another studyreports that hand skin temperature increases with pos-itive affect and decreases with threatening and unpleas-ant tasks (Rimm-Kaufman amp Kagan 1996)

RESULTS

Our sample of subjects consisted of 10 professionaltraders employed by the foreign-exchange and inter-est-rate derivatives business unit of a major globalfinancial institution based in Boston MA This institu-tion provides banking and other financial services toclients ranging from small regional startups to Fortune500 multinational corporations The foreign-exchangeand interest-rate derivatives business unit employs ap-proximately 90 professionals of which two-thirds spe-cializes in the trading of foreign exchange and relatedinstruments and one-third specializes in the trading ofinterest-rate derivative securities In a typical day thisbusiness unit engages in 1000- 1200 trades with anaverage size of US$3- 5 million per trade Approximately80 of the trades are executed on behalf of the clientsof the financial institution with the remaining 20motivated by the financial institutionrsquos market-makingactivities

For each of the 10 subjects five physiological variableswere monitored in real time during the entire duration ofeach session while the subjects sat at their trading con-soles (see Figure 1) Six sensors were used SCR BVPbody temperature (TMP) respiration (RSP) and twoelectromyographic sensors (facial and forearm EMG)

The benefits of acquiring real-time physiological meas-urements in vivo must be balanced against the cost ofmeasurement error spurious signals and other statisticalartifacts which if left untreated can obscure and con-found any genuine signals in the data To eliminate asmany artifacts as possible while maximizing the informa-tional content of the data each of the five physiological

Lo and Repin 325

variables were preprocessed with filters calibrated toeach variable individually and adaptive time-windowswere used in place of fixed-length windows to matchthe event-driven nature of the market variables

After preprocessing the following seven features wereextracted from SCR BVP temperature and respirationsignals

1 times of onset of SCR responses2 amplitudes of SCR responses3 average HR (every 3 sec)4 BVP signal amplitudes5 respiration rates (every 5 sec)6 respiration amplitudes7 temperature changes (from the 10-sec lag)

These features were selected to reflect the differentproperties and different information contained in theraw physiological variables SCRs were characterized bysmooth lsquolsquobumpsrsquorsquo with a relatively fast onset and slowdecay hence the number of such bumps and theirrelative strength are excellent summary measures ofthe SCR signal Intervals of 3 and 5 sec for heart andrespiration rates were chosen as a compromise betweenthe need for finer time slices to distinguish the onsetand completion of physiological and market events andthe necessity of a sufficient number of heartbeats andrespiration cycles within each interval to compute anaverage Under normal conditions a typical subject willexhibit an average of three to four heartbeats per 3-secinterval and approximately three breaths per 5-sec in-terval Temperature was the slowest-varying physiolog-ical signal and the 10-sec interval to register its changereflected the maximum possible interval in our analysis

For each session real-time market data for key finan-cial instruments actively traded or monitored by thesubject were collected synchronously with the physio-logical data In particular across the 10 subjects a total of15 instruments were consideredmdash13 foreign currencies

and two futures contracts the euro (EUR) the Japaneseyen (JPY) the British pound (GBP) the Canadian dollar(CAD) the Swiss franc (CHF) the Australian dollar(AUD) the New Zealand dollar (NZD) the Mexican peso(MXP) the Brazilian real (BRL) the Argentinian peso(ARS) the Colombian peso (COP) the Venezuelan boli-var (VEB) the Chilean peso (CLP) SampP 500 futures(SPU) and eurodollar futures (EDU) For each securityfour time series were monitored (1) the bid price Pt

b (2)the ask price Pt

a (3) the bidask spread Xt sup2 Ptb iexcl Pt

aand (4) the net return Rt sup2 (Pt iexcl Ptiexcl1) Ptiexcl1 where Pt

denotes the mid-spread price at time t and all priceswere sampled every second

For each of the financial time series we identifiedthree classes of market events deviations trend-rever-sals and volatility events These events are often cited bytraders as significant developments that require height-ened attention potentially signaling a shift in marketdynamics and risk exposures With these definitions andcalibrations in hand we implemented an automaticprocedure for detecting deviations trend-reversals andvolatility events for all the relevant time series in eachsession Table 1 reports event counts for the financialtime series monitored for each subject Feature vectorswere then constructed for all detected market events inall time series for all subjects and a set of control featurevectors was generated by applying the same feature-extraction process to randomly selected windows con-taining no events of any kind One set of control featurevectors was constructed for deviation and trend-reversalevents (10-sec intervals) and a second set was con-structed for volatility-type events (5-min intervals) Atwo-sided t test was applied to each component of eachfeature vector and the corresponding control vector totest the null hypothesis that the feature vectors werestatistically indistinguishable from the control featurevectors

For volatility events which by definition occurred in5-min intervals (event windows) we constructed featurevectors containing the following information for theduration of the event window

1 number of SCR responses2 average SCR amplitude3 average HR4 average ratio of the BVP amplitude to local baseline5 average ratio of the BVP amplitude to global

baseline6 number of temperature changes exceeding 018F7 average respiration rate8 average respiration amplitude

where the local baseline for the BVP signal is the averagelevel during the event window and the global baseline isthe average over the entire recording session for eachtrader Control feature vectors were constructed byapplying the feature-extraction process to randomlyselected 5-min windows containing no events of any

Laptop with data acquisition software

Fiber optic cable Control unit

Market information (Reuters Bloomberg) Order screen etc TIBCO market spreadsheet

eal-time data feed the market data

acquisition computer

Figure 1 Typical set-up for the measurement of real-time physiolo-gical responses of financial traders during live trading sessions Real-time market data used by the traders are recorded synchronously andsubsequently analyzed together with the physiological response data

326 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

kind For the other two types of market eventsmdashdevia-tions and trend-reversalsmdashlsquolsquopre-eventrsquorsquo and lsquolsquopost-eventrsquorsquofeature vectors were constructed before and after eachmarket event using the same variables where featureswere aggregated over 10-sec windows immediately pre-ceding and following the event Control feature vectorswere generated for these cases as well

The motivation for the post-event feature vector wasto capture the subjectrsquos reaction to the event with thepre-event feature vector as a benchmark from which tomeasure the magnitude of the reaction A comparisonbetween the pre-event feature vector and a controlfeature vector may provide an indication of a subjectrsquosanticipation of the event Latencies of the autonomicresponses reported in previous studies were on a timescale of 1 to 10 sec (see Cacioppo et al 2000) hence10-sec event windows were judged to be long enough forevent-related autonomic responses to occur and at thesame time short enough to minimize the likelihood ofoverlaps with other events or anomalies Five-minutewindows were used for volatility events because 10-secintervals were simply insufficient for meaningful volatilitycalculations For all of the financial time series used inthis study and for most financial time series in generalthere are few volatility events that occur in any 10-secinterval except of course under extreme conditions forexample the stock market crash of October 19 1987 (nosuch conditions prevailed during any of our sessions)

The statistical analysis of the physiological and marketdata was motivated by four objectives (1) to identifyparticular classes of events with statistically significantdifferences in mean autonomic responses before or after

an event as compared to the no-event control (2) toidentify particular traders or groups of traders based onexperience or other personal characteristics that dem-onstrate significant mean autonomic responses duringmarket events (3) to identify particular financial instru-ments or groups of instruments that are associated withsignificant mean autonomic responses and (4) to iden-tify particular physiological variables that demonstratesignificant mean response levels immediately followingthe event or during the event-anticipation period ascompared to the no-event control

To address the first objective a total of eight types ofevents were used for each of the time series

1 price deviations2 spread deviations3 return deviations4 price trend-reversals5 spread trend-reversals6 maximum volatility7 price volatility8 return volatility

and for each type of event a two-sided t test wasperformed for the pooled sample of all subjects and alltime series in which mean autonomic responses duringevent periods were compared to mean autonomic re-sponses during no-event control periods

The t statistics corresponding to the two-sided hypoth-esis test are given in the left subpanel of Table 2 labelledlsquolsquoAll Tradersrsquorsquo Statistics that are significant at the 5levelmdashbased on the t distribution with the appropriatedegrees of freedom for each entrymdashare highlighted For

Table 1 Number of Deviation (DEV) Trend-Reversal (TRV) and Volatility (VOL) Events Detected in Real-Time Market Data forEach Trader Over the Course of Each Trading Session

EUR JPYOther MajorCu rrenciesa

Latin Am ericanCu rrenciesb Derivativesc

Trader ID DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL

B32 21 16 15 25 17 15 - - - - - - - - -

B34 18 10 14 17 17 14 - - - - - - - - -

B35 20 16 14 13 17 15 21 9 15 24 16 13 - - -

B36 - - - - - - - - - - - - 28 23 31

B37 19 19 14 18 18 12 18 18 14 105 86 63 - - -

B38 21 20 12 22 18 13 23 20 15 96 61 56 - - -

B39 - - - - - - - - - - - - 40 37 23

B310 10 17 10 18 16 14 116 59 61 - - - - - -

B311 17 15 15 17 16 15 94 61 67 - - - - - -

B312 19 17 15 16 19 13 107 83 71 - - - - - -

aGBP CAD CHF AUD NZDbMXP BRL ARS COP VEB CLPcEDU SPU

Lo and Repin 327

Tab

le2

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vec

tors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

All

Tra

ders

Hig

hEx

peri

ence

Low

orM

oder

ate

Expe

rien

ce

Phys

iolo

gy

Fea

ture

s

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VOL

Ma

xim

um

DE

V

Pri

ce

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

Nu

mbe

ro

fSC

Rre

spon

ses

28

72

060

42

08

81

66

40

636

057

90

561

122

8iexcl

022

1iexcl

126

10

184

20

81

26

96

012

32

75

22

17

0iexcl

066

5iexcl

068

8

Aver

age

SCR

ampl

itud

eiexcl

069

50

887

018

6iexcl

247

81

136

088

00

030

034

40

744

iexcl1

340

114

90

904

075

61

578

iexcl1

265

iexcl1

947

iexcl0

012

iexcl0

615

Aver

age

HR

090

5iexcl

058

8iexcl

059

6iexcl

061

71

032

iexcl0

270

044

91

88

6iexcl

179

7iexcl

163

51

83

6iexcl

201

81

310

005

1iexcl

038

2iexcl

110

6iexcl

031

11

283

Aver

age

BV

Pam

plit

ude

rela

tive

tolo

cal

base

line

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9iexcl

006

1iexcl

074

70

334

001

50

164

26

23

iexcl0

204

iexcl1

775

087

51

79

6iexcl

128

92

70

8

Aver

age

BV

Pam

plit

ude

rela

tive

togl

obal

base

line

141

7iexcl

272

62

32

41

417

iexcl2

211

28

86

iexcl2

529

iexcl1

367

159

01

110

iexcl1

165

23

01

19

17

iexcl2

664

18

84

20

82

iexcl1

759

27

58

Nu

mbe

ro

fte

mp

erat

ure

jum

ps

098

82

82

00

650

17

77

48

68

iexcl2

482

iexcl1

010

iexcl1

435

NA

NA

NA

164

40

837

29

02

114

52

44

52

63

3iexcl

321

3

Aver

age

resp

irat

ion

rate

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4iexcl

155

61

109

012

10

866

iexcl0

102

009

30

079

iexcl0

552

032

3iexcl

063

1iexcl

217

0iexcl

152

8

Aver

age

resp

irat

ion

ampl

itud

e

iexcl1

646

iexcl0

250

013

60

034

iexcl1

487

iexcl3

499

108

70

265

iexcl1

777

iexcl0

657

012

1iexcl

060

7iexcl

274

9iexcl

131

92

14

20

055

iexcl2

832

iexcl4

533

Bo

lded

stat

isti

csar

esi

gnifi

can

tat

the

5le

vel

Th

ele

ftp

anel

give

sag

greg

ate

resu

lts

for

allt

rad

ers

Sim

ilar

ttes

tsfo

rtr

ader

sw

ith

hig

hex

per

ienc

ean

dlo

wo

rm

oder

ate

exp

erie

nce

are

show

nin

the

mid

dle

and

righ

tp

anel

sre

spec

tive

lyE

ntr

ies

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

wer

ep

rese

nt

inei

ther

the

even

tor

cont

rol

feat

ure

vect

or

for

the

par

ticu

lar

mar

ket-

even

tty

pe

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spo

nse

sH

R=

hea

rtra

te

BV

P=

blo

od

volu

me

pul

se

328 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

deviations and trend-reversals both the number of SCRsand the number of temperature jumps reached statisti-cally significant levels for three out of the five event typesBVP amplitude-related features were highly significantfor volatility eventsmdashsignificance levels less than 1 wereobtained for both BVP amplitude features for all threetypes of volatility events Because the results were sosimilar across the three types of volatility events (max-imum volatility price volatility and return volatility) wereport the results only for maximum volatility in Table 2Such patterns of autonomic responses may indicate thepresence of transient emotional responses to deviationsand trend-reversal events that occur within 10-sec win-dows while volatility events defined in 5-min windowselicited a tonic response in BVP

To explore potential differences between experi-enced and less experienced traders a sample of 10traders was divided into two groups each consistingof five traders the first containing highly experiencedtraders and the second containing traders with low ormoderate experience where the levels of experiencewere determined through interviews with the tradersrsquosupervisors The results for each of the two groupsare reported in the two right subpanels of Table 2labelled lsquolsquoHigh Experiencersquorsquo and lsquolsquoLow or ModerateExperiencersquorsquo The high-experience traders generallyexhibited low responses for deviations and trend-re-versal events with only two types of market events(spread deviations and spread trend-reversals) elicitingsignificant mean autonomic responses (average HR)However even for high-experience traders volatilityevents induced statistically significant mean responsesfor both SCR and BVP amplitude Less experiencedtraders showed a much higher number of significantmean responses in the number of SCRs BVP ampli-tude and number of temperature increases for devia-tions and trend-reversal events Table 2 shows thatsessions with low- and moderate-experience tradersyielded 13 t statistics significant at the 5 level whilesessions with high-experience traders yielded only fivet statistics significant at the 5 level For both sets oftraders volatility events were associated with signifi-cant BVP amplitude The differences in responsepatterns for deviations and trend-reversals observedin this case indicate that less experienced traders maybe more sensitive to short-term changes in the marketvariables than their more experienced colleagues

To address the third objective 10-sec intervals imme-diately preceding and immediately following each devia-tion and trend-reversal event were compared to control(ie no-event) time intervals Two separate t tests wereconductedmdashone for pre- and another for post-eventtime intervalsmdashfor each of the three types of deviationsand two types of trend-reversals Because the objectivewas to detect differences in how pre- and post-eventfeature vectors differed from the control we used thesame control feature vectors for both sets of t tests In all

cases the t tests were designed to test the null hypoth-esis that both pre- and post-event feature vectors werestatistically indistinguishable from the control featurevector Surprisingly these two sets of t tests yielded verysimilar results reported in Table 3 implying that noneof the physiological variables were predictors of antici-patory emotional responses These findings may be atleast partly explained by how the events were defined Inparticular the definitions of deviation and trend-reversalevents are those instances where the time seriesachieved a prespecified deviation from the time-seriesmean and its moving average respectively In theabsence of large jumps the values of the time seriesnear an event defined in this way are likely to becomparable to the event itself Therefore the tradersmay be responding to market conditions occurringthroughout the pre- and post-event period not just atthe exact time of the event hence the similarity be-tween the two periods More complex definitions ofevents may allow us to discriminate between pre- andpost-event physiological responses and we are explor-ing several alternatives in ongoing research

Finally to address the fourth objective the followingfour groups of financial instruments were formed on thebasis of similarity in their statistical characteristics

Group 1 EUR JPY

Group 2 GBP CAD CHF AUD NZD (other majorcurrencies)

Group 3 MXP BRL ARS COP VEB CLP (LatinAmerican currencies)

Group 4 SPU EDU (derivatives)

Group 1 is by far the most actively traded set offinancial instruments and accounts for most of thetrading activities of our sample of 10 traders hence itis not surprising that the pattern of statistical signifi-cance for Group 1 in Table 4 is almost identical to thepattern for all traders in Table 2 (the left panel) SCRis significant for price and return deviations temper-ature jumps are significant for spread and price devia-tions and both measures of BVP are significant forvolatility events For Group 2 none of the eightphysiological variables were significant for price orspread deviations although temperature changes andrespiration rates were significant for return deviationsand both BVP measures were significant for pricetrend-reversals and volatility events The financial in-struments in Group 2 were not the primary focus ofany of the traders in our sample which may explainthe different patterns of significance as compared withGroup 1 Group 4 contains the fewest significantresponses and this is consistent with the fact thatthese securities were traded by two of the mostexperienced traders in our sample Derivative secur-ities are considerably more complex instruments thanthe spot currencies requiring more skill and trainingthan the other groups of securities However SCR was

Lo and Repin 329

Tab

le3

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gPr

e-an

dPo

st-e

vent

Tim

eIn

terv

als

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

Pre

-eve

nt

Inte

rva

lP

ost-

even

tIn

terv

al

Phy

sio

logy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mbe

rof

SCR

resp

ons

es3

75

61

543

25

40

21

18

24

01

057

92

87

20

604

20

88

16

64

063

60

579

Aver

age

SCR

amp

litu

de0

700

iexcl0

454

iexcl1

793

051

11

068

088

0iexcl

069

50

887

018

6iexcl

247

81

136

088

0

Aver

age

HR

096

8iexcl

020

5iexcl

055

2iexcl

066

3iexcl

038

6iexcl

027

00

905

iexcl0

588

iexcl0

596

iexcl0

617

103

2iexcl

027

0

Aver

age

BVP

amp

litud

ere

lativ

eto

loca

lba

selin

e0

043

iexcl0

895

006

00

814

iexcl0

045

36

19

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9

Aver

age

BVP

amp

litud

ere

lativ

eto

glob

alba

selin

e1

183

iexcl2

794

23

97

131

7iexcl

225

62

88

61

417

iexcl2

726

23

24

141

7iexcl

221

12

88

6

Nu

mbe

rof

tem

per

atur

eju

mp

s1

110

27

67

050

52

15

83

92

5iexcl

248

20

988

28

20

065

01

77

74

86

8iexcl

248

2

Aver

age

resp

irat

ion

rate

126

1iexcl

165

0iexcl

025

41

303

iexcl1

999

iexcl0

064

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4

Aver

age

resp

irat

ion

amp

litud

eiexcl

162

0iexcl

006

70

561

006

8iexcl

180

7iexcl

349

9iexcl

164

6iexcl

025

00

136

003

4iexcl

148

7iexcl

349

9

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

l

Th

ele

ftp

anel

cont

ain

st

stat

isti

csfo

rp

re-e

vent

feat

ure

vect

ors

and

the

righ

tp

anel

con

tain

st

stat

isti

csfo

rp

ost-

even

tfe

atu

reve

cto

rs

both

test

edag

ain

stth

esa

me

set

of

con

tro

ls

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spon

ses

HR

=h

eart

rate

B

VP

=bl

ood

volu

me

pu

lse

330 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

variables were preprocessed with filters calibrated toeach variable individually and adaptive time-windowswere used in place of fixed-length windows to matchthe event-driven nature of the market variables

After preprocessing the following seven features wereextracted from SCR BVP temperature and respirationsignals

1 times of onset of SCR responses2 amplitudes of SCR responses3 average HR (every 3 sec)4 BVP signal amplitudes5 respiration rates (every 5 sec)6 respiration amplitudes7 temperature changes (from the 10-sec lag)

These features were selected to reflect the differentproperties and different information contained in theraw physiological variables SCRs were characterized bysmooth lsquolsquobumpsrsquorsquo with a relatively fast onset and slowdecay hence the number of such bumps and theirrelative strength are excellent summary measures ofthe SCR signal Intervals of 3 and 5 sec for heart andrespiration rates were chosen as a compromise betweenthe need for finer time slices to distinguish the onsetand completion of physiological and market events andthe necessity of a sufficient number of heartbeats andrespiration cycles within each interval to compute anaverage Under normal conditions a typical subject willexhibit an average of three to four heartbeats per 3-secinterval and approximately three breaths per 5-sec in-terval Temperature was the slowest-varying physiolog-ical signal and the 10-sec interval to register its changereflected the maximum possible interval in our analysis

For each session real-time market data for key finan-cial instruments actively traded or monitored by thesubject were collected synchronously with the physio-logical data In particular across the 10 subjects a total of15 instruments were consideredmdash13 foreign currencies

and two futures contracts the euro (EUR) the Japaneseyen (JPY) the British pound (GBP) the Canadian dollar(CAD) the Swiss franc (CHF) the Australian dollar(AUD) the New Zealand dollar (NZD) the Mexican peso(MXP) the Brazilian real (BRL) the Argentinian peso(ARS) the Colombian peso (COP) the Venezuelan boli-var (VEB) the Chilean peso (CLP) SampP 500 futures(SPU) and eurodollar futures (EDU) For each securityfour time series were monitored (1) the bid price Pt

b (2)the ask price Pt

a (3) the bidask spread Xt sup2 Ptb iexcl Pt

aand (4) the net return Rt sup2 (Pt iexcl Ptiexcl1) Ptiexcl1 where Pt

denotes the mid-spread price at time t and all priceswere sampled every second

For each of the financial time series we identifiedthree classes of market events deviations trend-rever-sals and volatility events These events are often cited bytraders as significant developments that require height-ened attention potentially signaling a shift in marketdynamics and risk exposures With these definitions andcalibrations in hand we implemented an automaticprocedure for detecting deviations trend-reversals andvolatility events for all the relevant time series in eachsession Table 1 reports event counts for the financialtime series monitored for each subject Feature vectorswere then constructed for all detected market events inall time series for all subjects and a set of control featurevectors was generated by applying the same feature-extraction process to randomly selected windows con-taining no events of any kind One set of control featurevectors was constructed for deviation and trend-reversalevents (10-sec intervals) and a second set was con-structed for volatility-type events (5-min intervals) Atwo-sided t test was applied to each component of eachfeature vector and the corresponding control vector totest the null hypothesis that the feature vectors werestatistically indistinguishable from the control featurevectors

For volatility events which by definition occurred in5-min intervals (event windows) we constructed featurevectors containing the following information for theduration of the event window

1 number of SCR responses2 average SCR amplitude3 average HR4 average ratio of the BVP amplitude to local baseline5 average ratio of the BVP amplitude to global

baseline6 number of temperature changes exceeding 018F7 average respiration rate8 average respiration amplitude

where the local baseline for the BVP signal is the averagelevel during the event window and the global baseline isthe average over the entire recording session for eachtrader Control feature vectors were constructed byapplying the feature-extraction process to randomlyselected 5-min windows containing no events of any

Laptop with data acquisition software

Fiber optic cable Control unit

Market information (Reuters Bloomberg) Order screen etc TIBCO market spreadsheet

eal-time data feed the market data

acquisition computer

Figure 1 Typical set-up for the measurement of real-time physiolo-gical responses of financial traders during live trading sessions Real-time market data used by the traders are recorded synchronously andsubsequently analyzed together with the physiological response data

326 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

kind For the other two types of market eventsmdashdevia-tions and trend-reversalsmdashlsquolsquopre-eventrsquorsquo and lsquolsquopost-eventrsquorsquofeature vectors were constructed before and after eachmarket event using the same variables where featureswere aggregated over 10-sec windows immediately pre-ceding and following the event Control feature vectorswere generated for these cases as well

The motivation for the post-event feature vector wasto capture the subjectrsquos reaction to the event with thepre-event feature vector as a benchmark from which tomeasure the magnitude of the reaction A comparisonbetween the pre-event feature vector and a controlfeature vector may provide an indication of a subjectrsquosanticipation of the event Latencies of the autonomicresponses reported in previous studies were on a timescale of 1 to 10 sec (see Cacioppo et al 2000) hence10-sec event windows were judged to be long enough forevent-related autonomic responses to occur and at thesame time short enough to minimize the likelihood ofoverlaps with other events or anomalies Five-minutewindows were used for volatility events because 10-secintervals were simply insufficient for meaningful volatilitycalculations For all of the financial time series used inthis study and for most financial time series in generalthere are few volatility events that occur in any 10-secinterval except of course under extreme conditions forexample the stock market crash of October 19 1987 (nosuch conditions prevailed during any of our sessions)

The statistical analysis of the physiological and marketdata was motivated by four objectives (1) to identifyparticular classes of events with statistically significantdifferences in mean autonomic responses before or after

an event as compared to the no-event control (2) toidentify particular traders or groups of traders based onexperience or other personal characteristics that dem-onstrate significant mean autonomic responses duringmarket events (3) to identify particular financial instru-ments or groups of instruments that are associated withsignificant mean autonomic responses and (4) to iden-tify particular physiological variables that demonstratesignificant mean response levels immediately followingthe event or during the event-anticipation period ascompared to the no-event control

To address the first objective a total of eight types ofevents were used for each of the time series

1 price deviations2 spread deviations3 return deviations4 price trend-reversals5 spread trend-reversals6 maximum volatility7 price volatility8 return volatility

and for each type of event a two-sided t test wasperformed for the pooled sample of all subjects and alltime series in which mean autonomic responses duringevent periods were compared to mean autonomic re-sponses during no-event control periods

The t statistics corresponding to the two-sided hypoth-esis test are given in the left subpanel of Table 2 labelledlsquolsquoAll Tradersrsquorsquo Statistics that are significant at the 5levelmdashbased on the t distribution with the appropriatedegrees of freedom for each entrymdashare highlighted For

Table 1 Number of Deviation (DEV) Trend-Reversal (TRV) and Volatility (VOL) Events Detected in Real-Time Market Data forEach Trader Over the Course of Each Trading Session

EUR JPYOther MajorCu rrenciesa

Latin Am ericanCu rrenciesb Derivativesc

Trader ID DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL

B32 21 16 15 25 17 15 - - - - - - - - -

B34 18 10 14 17 17 14 - - - - - - - - -

B35 20 16 14 13 17 15 21 9 15 24 16 13 - - -

B36 - - - - - - - - - - - - 28 23 31

B37 19 19 14 18 18 12 18 18 14 105 86 63 - - -

B38 21 20 12 22 18 13 23 20 15 96 61 56 - - -

B39 - - - - - - - - - - - - 40 37 23

B310 10 17 10 18 16 14 116 59 61 - - - - - -

B311 17 15 15 17 16 15 94 61 67 - - - - - -

B312 19 17 15 16 19 13 107 83 71 - - - - - -

aGBP CAD CHF AUD NZDbMXP BRL ARS COP VEB CLPcEDU SPU

Lo and Repin 327

Tab

le2

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vec

tors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

All

Tra

ders

Hig

hEx

peri

ence

Low

orM

oder

ate

Expe

rien

ce

Phys

iolo

gy

Fea

ture

s

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VOL

Ma

xim

um

DE

V

Pri

ce

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

Nu

mbe

ro

fSC

Rre

spon

ses

28

72

060

42

08

81

66

40

636

057

90

561

122

8iexcl

022

1iexcl

126

10

184

20

81

26

96

012

32

75

22

17

0iexcl

066

5iexcl

068

8

Aver

age

SCR

ampl

itud

eiexcl

069

50

887

018

6iexcl

247

81

136

088

00

030

034

40

744

iexcl1

340

114

90

904

075

61

578

iexcl1

265

iexcl1

947

iexcl0

012

iexcl0

615

Aver

age

HR

090

5iexcl

058

8iexcl

059

6iexcl

061

71

032

iexcl0

270

044

91

88

6iexcl

179

7iexcl

163

51

83

6iexcl

201

81

310

005

1iexcl

038

2iexcl

110

6iexcl

031

11

283

Aver

age

BV

Pam

plit

ude

rela

tive

tolo

cal

base

line

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9iexcl

006

1iexcl

074

70

334

001

50

164

26

23

iexcl0

204

iexcl1

775

087

51

79

6iexcl

128

92

70

8

Aver

age

BV

Pam

plit

ude

rela

tive

togl

obal

base

line

141

7iexcl

272

62

32

41

417

iexcl2

211

28

86

iexcl2

529

iexcl1

367

159

01

110

iexcl1

165

23

01

19

17

iexcl2

664

18

84

20

82

iexcl1

759

27

58

Nu

mbe

ro

fte

mp

erat

ure

jum

ps

098

82

82

00

650

17

77

48

68

iexcl2

482

iexcl1

010

iexcl1

435

NA

NA

NA

164

40

837

29

02

114

52

44

52

63

3iexcl

321

3

Aver

age

resp

irat

ion

rate

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4iexcl

155

61

109

012

10

866

iexcl0

102

009

30

079

iexcl0

552

032

3iexcl

063

1iexcl

217

0iexcl

152

8

Aver

age

resp

irat

ion

ampl

itud

e

iexcl1

646

iexcl0

250

013

60

034

iexcl1

487

iexcl3

499

108

70

265

iexcl1

777

iexcl0

657

012

1iexcl

060

7iexcl

274

9iexcl

131

92

14

20

055

iexcl2

832

iexcl4

533

Bo

lded

stat

isti

csar

esi

gnifi

can

tat

the

5le

vel

Th

ele

ftp

anel

give

sag

greg

ate

resu

lts

for

allt

rad

ers

Sim

ilar

ttes

tsfo

rtr

ader

sw

ith

hig

hex

per

ienc

ean

dlo

wo

rm

oder

ate

exp

erie

nce

are

show

nin

the

mid

dle

and

righ

tp

anel

sre

spec

tive

lyE

ntr

ies

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

wer

ep

rese

nt

inei

ther

the

even

tor

cont

rol

feat

ure

vect

or

for

the

par

ticu

lar

mar

ket-

even

tty

pe

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spo

nse

sH

R=

hea

rtra

te

BV

P=

blo

od

volu

me

pul

se

328 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

deviations and trend-reversals both the number of SCRsand the number of temperature jumps reached statisti-cally significant levels for three out of the five event typesBVP amplitude-related features were highly significantfor volatility eventsmdashsignificance levels less than 1 wereobtained for both BVP amplitude features for all threetypes of volatility events Because the results were sosimilar across the three types of volatility events (max-imum volatility price volatility and return volatility) wereport the results only for maximum volatility in Table 2Such patterns of autonomic responses may indicate thepresence of transient emotional responses to deviationsand trend-reversal events that occur within 10-sec win-dows while volatility events defined in 5-min windowselicited a tonic response in BVP

To explore potential differences between experi-enced and less experienced traders a sample of 10traders was divided into two groups each consistingof five traders the first containing highly experiencedtraders and the second containing traders with low ormoderate experience where the levels of experiencewere determined through interviews with the tradersrsquosupervisors The results for each of the two groupsare reported in the two right subpanels of Table 2labelled lsquolsquoHigh Experiencersquorsquo and lsquolsquoLow or ModerateExperiencersquorsquo The high-experience traders generallyexhibited low responses for deviations and trend-re-versal events with only two types of market events(spread deviations and spread trend-reversals) elicitingsignificant mean autonomic responses (average HR)However even for high-experience traders volatilityevents induced statistically significant mean responsesfor both SCR and BVP amplitude Less experiencedtraders showed a much higher number of significantmean responses in the number of SCRs BVP ampli-tude and number of temperature increases for devia-tions and trend-reversal events Table 2 shows thatsessions with low- and moderate-experience tradersyielded 13 t statistics significant at the 5 level whilesessions with high-experience traders yielded only fivet statistics significant at the 5 level For both sets oftraders volatility events were associated with signifi-cant BVP amplitude The differences in responsepatterns for deviations and trend-reversals observedin this case indicate that less experienced traders maybe more sensitive to short-term changes in the marketvariables than their more experienced colleagues

To address the third objective 10-sec intervals imme-diately preceding and immediately following each devia-tion and trend-reversal event were compared to control(ie no-event) time intervals Two separate t tests wereconductedmdashone for pre- and another for post-eventtime intervalsmdashfor each of the three types of deviationsand two types of trend-reversals Because the objectivewas to detect differences in how pre- and post-eventfeature vectors differed from the control we used thesame control feature vectors for both sets of t tests In all

cases the t tests were designed to test the null hypoth-esis that both pre- and post-event feature vectors werestatistically indistinguishable from the control featurevector Surprisingly these two sets of t tests yielded verysimilar results reported in Table 3 implying that noneof the physiological variables were predictors of antici-patory emotional responses These findings may be atleast partly explained by how the events were defined Inparticular the definitions of deviation and trend-reversalevents are those instances where the time seriesachieved a prespecified deviation from the time-seriesmean and its moving average respectively In theabsence of large jumps the values of the time seriesnear an event defined in this way are likely to becomparable to the event itself Therefore the tradersmay be responding to market conditions occurringthroughout the pre- and post-event period not just atthe exact time of the event hence the similarity be-tween the two periods More complex definitions ofevents may allow us to discriminate between pre- andpost-event physiological responses and we are explor-ing several alternatives in ongoing research

Finally to address the fourth objective the followingfour groups of financial instruments were formed on thebasis of similarity in their statistical characteristics

Group 1 EUR JPY

Group 2 GBP CAD CHF AUD NZD (other majorcurrencies)

Group 3 MXP BRL ARS COP VEB CLP (LatinAmerican currencies)

Group 4 SPU EDU (derivatives)

Group 1 is by far the most actively traded set offinancial instruments and accounts for most of thetrading activities of our sample of 10 traders hence itis not surprising that the pattern of statistical signifi-cance for Group 1 in Table 4 is almost identical to thepattern for all traders in Table 2 (the left panel) SCRis significant for price and return deviations temper-ature jumps are significant for spread and price devia-tions and both measures of BVP are significant forvolatility events For Group 2 none of the eightphysiological variables were significant for price orspread deviations although temperature changes andrespiration rates were significant for return deviationsand both BVP measures were significant for pricetrend-reversals and volatility events The financial in-struments in Group 2 were not the primary focus ofany of the traders in our sample which may explainthe different patterns of significance as compared withGroup 1 Group 4 contains the fewest significantresponses and this is consistent with the fact thatthese securities were traded by two of the mostexperienced traders in our sample Derivative secur-ities are considerably more complex instruments thanthe spot currencies requiring more skill and trainingthan the other groups of securities However SCR was

Lo and Repin 329

Tab

le3

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gPr

e-an

dPo

st-e

vent

Tim

eIn

terv

als

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

Pre

-eve

nt

Inte

rva

lP

ost-

even

tIn

terv

al

Phy

sio

logy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mbe

rof

SCR

resp

ons

es3

75

61

543

25

40

21

18

24

01

057

92

87

20

604

20

88

16

64

063

60

579

Aver

age

SCR

amp

litu

de0

700

iexcl0

454

iexcl1

793

051

11

068

088

0iexcl

069

50

887

018

6iexcl

247

81

136

088

0

Aver

age

HR

096

8iexcl

020

5iexcl

055

2iexcl

066

3iexcl

038

6iexcl

027

00

905

iexcl0

588

iexcl0

596

iexcl0

617

103

2iexcl

027

0

Aver

age

BVP

amp

litud

ere

lativ

eto

loca

lba

selin

e0

043

iexcl0

895

006

00

814

iexcl0

045

36

19

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9

Aver

age

BVP

amp

litud

ere

lativ

eto

glob

alba

selin

e1

183

iexcl2

794

23

97

131

7iexcl

225

62

88

61

417

iexcl2

726

23

24

141

7iexcl

221

12

88

6

Nu

mbe

rof

tem

per

atur

eju

mp

s1

110

27

67

050

52

15

83

92

5iexcl

248

20

988

28

20

065

01

77

74

86

8iexcl

248

2

Aver

age

resp

irat

ion

rate

126

1iexcl

165

0iexcl

025

41

303

iexcl1

999

iexcl0

064

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4

Aver

age

resp

irat

ion

amp

litud

eiexcl

162

0iexcl

006

70

561

006

8iexcl

180

7iexcl

349

9iexcl

164

6iexcl

025

00

136

003

4iexcl

148

7iexcl

349

9

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

l

Th

ele

ftp

anel

cont

ain

st

stat

isti

csfo

rp

re-e

vent

feat

ure

vect

ors

and

the

righ

tp

anel

con

tain

st

stat

isti

csfo

rp

ost-

even

tfe

atu

reve

cto

rs

both

test

edag

ain

stth

esa

me

set

of

con

tro

ls

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spon

ses

HR

=h

eart

rate

B

VP

=bl

ood

volu

me

pu

lse

330 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

kind For the other two types of market eventsmdashdevia-tions and trend-reversalsmdashlsquolsquopre-eventrsquorsquo and lsquolsquopost-eventrsquorsquofeature vectors were constructed before and after eachmarket event using the same variables where featureswere aggregated over 10-sec windows immediately pre-ceding and following the event Control feature vectorswere generated for these cases as well

The motivation for the post-event feature vector wasto capture the subjectrsquos reaction to the event with thepre-event feature vector as a benchmark from which tomeasure the magnitude of the reaction A comparisonbetween the pre-event feature vector and a controlfeature vector may provide an indication of a subjectrsquosanticipation of the event Latencies of the autonomicresponses reported in previous studies were on a timescale of 1 to 10 sec (see Cacioppo et al 2000) hence10-sec event windows were judged to be long enough forevent-related autonomic responses to occur and at thesame time short enough to minimize the likelihood ofoverlaps with other events or anomalies Five-minutewindows were used for volatility events because 10-secintervals were simply insufficient for meaningful volatilitycalculations For all of the financial time series used inthis study and for most financial time series in generalthere are few volatility events that occur in any 10-secinterval except of course under extreme conditions forexample the stock market crash of October 19 1987 (nosuch conditions prevailed during any of our sessions)

The statistical analysis of the physiological and marketdata was motivated by four objectives (1) to identifyparticular classes of events with statistically significantdifferences in mean autonomic responses before or after

an event as compared to the no-event control (2) toidentify particular traders or groups of traders based onexperience or other personal characteristics that dem-onstrate significant mean autonomic responses duringmarket events (3) to identify particular financial instru-ments or groups of instruments that are associated withsignificant mean autonomic responses and (4) to iden-tify particular physiological variables that demonstratesignificant mean response levels immediately followingthe event or during the event-anticipation period ascompared to the no-event control

To address the first objective a total of eight types ofevents were used for each of the time series

1 price deviations2 spread deviations3 return deviations4 price trend-reversals5 spread trend-reversals6 maximum volatility7 price volatility8 return volatility

and for each type of event a two-sided t test wasperformed for the pooled sample of all subjects and alltime series in which mean autonomic responses duringevent periods were compared to mean autonomic re-sponses during no-event control periods

The t statistics corresponding to the two-sided hypoth-esis test are given in the left subpanel of Table 2 labelledlsquolsquoAll Tradersrsquorsquo Statistics that are significant at the 5levelmdashbased on the t distribution with the appropriatedegrees of freedom for each entrymdashare highlighted For

Table 1 Number of Deviation (DEV) Trend-Reversal (TRV) and Volatility (VOL) Events Detected in Real-Time Market Data forEach Trader Over the Course of Each Trading Session

EUR JPYOther MajorCu rrenciesa

Latin Am ericanCu rrenciesb Derivativesc

Trader ID DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL DEV TRV VOL

B32 21 16 15 25 17 15 - - - - - - - - -

B34 18 10 14 17 17 14 - - - - - - - - -

B35 20 16 14 13 17 15 21 9 15 24 16 13 - - -

B36 - - - - - - - - - - - - 28 23 31

B37 19 19 14 18 18 12 18 18 14 105 86 63 - - -

B38 21 20 12 22 18 13 23 20 15 96 61 56 - - -

B39 - - - - - - - - - - - - 40 37 23

B310 10 17 10 18 16 14 116 59 61 - - - - - -

B311 17 15 15 17 16 15 94 61 67 - - - - - -

B312 19 17 15 16 19 13 107 83 71 - - - - - -

aGBP CAD CHF AUD NZDbMXP BRL ARS COP VEB CLPcEDU SPU

Lo and Repin 327

Tab

le2

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vec

tors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

All

Tra

ders

Hig

hEx

peri

ence

Low

orM

oder

ate

Expe

rien

ce

Phys

iolo

gy

Fea

ture

s

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VOL

Ma

xim

um

DE

V

Pri

ce

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

Nu

mbe

ro

fSC

Rre

spon

ses

28

72

060

42

08

81

66

40

636

057

90

561

122

8iexcl

022

1iexcl

126

10

184

20

81

26

96

012

32

75

22

17

0iexcl

066

5iexcl

068

8

Aver

age

SCR

ampl

itud

eiexcl

069

50

887

018

6iexcl

247

81

136

088

00

030

034

40

744

iexcl1

340

114

90

904

075

61

578

iexcl1

265

iexcl1

947

iexcl0

012

iexcl0

615

Aver

age

HR

090

5iexcl

058

8iexcl

059

6iexcl

061

71

032

iexcl0

270

044

91

88

6iexcl

179

7iexcl

163

51

83

6iexcl

201

81

310

005

1iexcl

038

2iexcl

110

6iexcl

031

11

283

Aver

age

BV

Pam

plit

ude

rela

tive

tolo

cal

base

line

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9iexcl

006

1iexcl

074

70

334

001

50

164

26

23

iexcl0

204

iexcl1

775

087

51

79

6iexcl

128

92

70

8

Aver

age

BV

Pam

plit

ude

rela

tive

togl

obal

base

line

141

7iexcl

272

62

32

41

417

iexcl2

211

28

86

iexcl2

529

iexcl1

367

159

01

110

iexcl1

165

23

01

19

17

iexcl2

664

18

84

20

82

iexcl1

759

27

58

Nu

mbe

ro

fte

mp

erat

ure

jum

ps

098

82

82

00

650

17

77

48

68

iexcl2

482

iexcl1

010

iexcl1

435

NA

NA

NA

164

40

837

29

02

114

52

44

52

63

3iexcl

321

3

Aver

age

resp

irat

ion

rate

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4iexcl

155

61

109

012

10

866

iexcl0

102

009

30

079

iexcl0

552

032

3iexcl

063

1iexcl

217

0iexcl

152

8

Aver

age

resp

irat

ion

ampl

itud

e

iexcl1

646

iexcl0

250

013

60

034

iexcl1

487

iexcl3

499

108

70

265

iexcl1

777

iexcl0

657

012

1iexcl

060

7iexcl

274

9iexcl

131

92

14

20

055

iexcl2

832

iexcl4

533

Bo

lded

stat

isti

csar

esi

gnifi

can

tat

the

5le

vel

Th

ele

ftp

anel

give

sag

greg

ate

resu

lts

for

allt

rad

ers

Sim

ilar

ttes

tsfo

rtr

ader

sw

ith

hig

hex

per

ienc

ean

dlo

wo

rm

oder

ate

exp

erie

nce

are

show

nin

the

mid

dle

and

righ

tp

anel

sre

spec

tive

lyE

ntr

ies

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

wer

ep

rese

nt

inei

ther

the

even

tor

cont

rol

feat

ure

vect

or

for

the

par

ticu

lar

mar

ket-

even

tty

pe

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spo

nse

sH

R=

hea

rtra

te

BV

P=

blo

od

volu

me

pul

se

328 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

deviations and trend-reversals both the number of SCRsand the number of temperature jumps reached statisti-cally significant levels for three out of the five event typesBVP amplitude-related features were highly significantfor volatility eventsmdashsignificance levels less than 1 wereobtained for both BVP amplitude features for all threetypes of volatility events Because the results were sosimilar across the three types of volatility events (max-imum volatility price volatility and return volatility) wereport the results only for maximum volatility in Table 2Such patterns of autonomic responses may indicate thepresence of transient emotional responses to deviationsand trend-reversal events that occur within 10-sec win-dows while volatility events defined in 5-min windowselicited a tonic response in BVP

To explore potential differences between experi-enced and less experienced traders a sample of 10traders was divided into two groups each consistingof five traders the first containing highly experiencedtraders and the second containing traders with low ormoderate experience where the levels of experiencewere determined through interviews with the tradersrsquosupervisors The results for each of the two groupsare reported in the two right subpanels of Table 2labelled lsquolsquoHigh Experiencersquorsquo and lsquolsquoLow or ModerateExperiencersquorsquo The high-experience traders generallyexhibited low responses for deviations and trend-re-versal events with only two types of market events(spread deviations and spread trend-reversals) elicitingsignificant mean autonomic responses (average HR)However even for high-experience traders volatilityevents induced statistically significant mean responsesfor both SCR and BVP amplitude Less experiencedtraders showed a much higher number of significantmean responses in the number of SCRs BVP ampli-tude and number of temperature increases for devia-tions and trend-reversal events Table 2 shows thatsessions with low- and moderate-experience tradersyielded 13 t statistics significant at the 5 level whilesessions with high-experience traders yielded only fivet statistics significant at the 5 level For both sets oftraders volatility events were associated with signifi-cant BVP amplitude The differences in responsepatterns for deviations and trend-reversals observedin this case indicate that less experienced traders maybe more sensitive to short-term changes in the marketvariables than their more experienced colleagues

To address the third objective 10-sec intervals imme-diately preceding and immediately following each devia-tion and trend-reversal event were compared to control(ie no-event) time intervals Two separate t tests wereconductedmdashone for pre- and another for post-eventtime intervalsmdashfor each of the three types of deviationsand two types of trend-reversals Because the objectivewas to detect differences in how pre- and post-eventfeature vectors differed from the control we used thesame control feature vectors for both sets of t tests In all

cases the t tests were designed to test the null hypoth-esis that both pre- and post-event feature vectors werestatistically indistinguishable from the control featurevector Surprisingly these two sets of t tests yielded verysimilar results reported in Table 3 implying that noneof the physiological variables were predictors of antici-patory emotional responses These findings may be atleast partly explained by how the events were defined Inparticular the definitions of deviation and trend-reversalevents are those instances where the time seriesachieved a prespecified deviation from the time-seriesmean and its moving average respectively In theabsence of large jumps the values of the time seriesnear an event defined in this way are likely to becomparable to the event itself Therefore the tradersmay be responding to market conditions occurringthroughout the pre- and post-event period not just atthe exact time of the event hence the similarity be-tween the two periods More complex definitions ofevents may allow us to discriminate between pre- andpost-event physiological responses and we are explor-ing several alternatives in ongoing research

Finally to address the fourth objective the followingfour groups of financial instruments were formed on thebasis of similarity in their statistical characteristics

Group 1 EUR JPY

Group 2 GBP CAD CHF AUD NZD (other majorcurrencies)

Group 3 MXP BRL ARS COP VEB CLP (LatinAmerican currencies)

Group 4 SPU EDU (derivatives)

Group 1 is by far the most actively traded set offinancial instruments and accounts for most of thetrading activities of our sample of 10 traders hence itis not surprising that the pattern of statistical signifi-cance for Group 1 in Table 4 is almost identical to thepattern for all traders in Table 2 (the left panel) SCRis significant for price and return deviations temper-ature jumps are significant for spread and price devia-tions and both measures of BVP are significant forvolatility events For Group 2 none of the eightphysiological variables were significant for price orspread deviations although temperature changes andrespiration rates were significant for return deviationsand both BVP measures were significant for pricetrend-reversals and volatility events The financial in-struments in Group 2 were not the primary focus ofany of the traders in our sample which may explainthe different patterns of significance as compared withGroup 1 Group 4 contains the fewest significantresponses and this is consistent with the fact thatthese securities were traded by two of the mostexperienced traders in our sample Derivative secur-ities are considerably more complex instruments thanthe spot currencies requiring more skill and trainingthan the other groups of securities However SCR was

Lo and Repin 329

Tab

le3

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gPr

e-an

dPo

st-e

vent

Tim

eIn

terv

als

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

Pre

-eve

nt

Inte

rva

lP

ost-

even

tIn

terv

al

Phy

sio

logy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mbe

rof

SCR

resp

ons

es3

75

61

543

25

40

21

18

24

01

057

92

87

20

604

20

88

16

64

063

60

579

Aver

age

SCR

amp

litu

de0

700

iexcl0

454

iexcl1

793

051

11

068

088

0iexcl

069

50

887

018

6iexcl

247

81

136

088

0

Aver

age

HR

096

8iexcl

020

5iexcl

055

2iexcl

066

3iexcl

038

6iexcl

027

00

905

iexcl0

588

iexcl0

596

iexcl0

617

103

2iexcl

027

0

Aver

age

BVP

amp

litud

ere

lativ

eto

loca

lba

selin

e0

043

iexcl0

895

006

00

814

iexcl0

045

36

19

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9

Aver

age

BVP

amp

litud

ere

lativ

eto

glob

alba

selin

e1

183

iexcl2

794

23

97

131

7iexcl

225

62

88

61

417

iexcl2

726

23

24

141

7iexcl

221

12

88

6

Nu

mbe

rof

tem

per

atur

eju

mp

s1

110

27

67

050

52

15

83

92

5iexcl

248

20

988

28

20

065

01

77

74

86

8iexcl

248

2

Aver

age

resp

irat

ion

rate

126

1iexcl

165

0iexcl

025

41

303

iexcl1

999

iexcl0

064

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4

Aver

age

resp

irat

ion

amp

litud

eiexcl

162

0iexcl

006

70

561

006

8iexcl

180

7iexcl

349

9iexcl

164

6iexcl

025

00

136

003

4iexcl

148

7iexcl

349

9

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

l

Th

ele

ftp

anel

cont

ain

st

stat

isti

csfo

rp

re-e

vent

feat

ure

vect

ors

and

the

righ

tp

anel

con

tain

st

stat

isti

csfo

rp

ost-

even

tfe

atu

reve

cto

rs

both

test

edag

ain

stth

esa

me

set

of

con

tro

ls

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spon

ses

HR

=h

eart

rate

B

VP

=bl

ood

volu

me

pu

lse

330 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le2

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vec

tors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

All

Tra

ders

Hig

hEx

peri

ence

Low

orM

oder

ate

Expe

rien

ce

Phys

iolo

gy

Fea

ture

s

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VOL

Ma

xim

um

DE

V

Pri

ce

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

DE

V

Pri

ce

DEV

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ce

TRV

Spre

ad

VO

L

Ma

xim

um

Nu

mbe

ro

fSC

Rre

spon

ses

28

72

060

42

08

81

66

40

636

057

90

561

122

8iexcl

022

1iexcl

126

10

184

20

81

26

96

012

32

75

22

17

0iexcl

066

5iexcl

068

8

Aver

age

SCR

ampl

itud

eiexcl

069

50

887

018

6iexcl

247

81

136

088

00

030

034

40

744

iexcl1

340

114

90

904

075

61

578

iexcl1

265

iexcl1

947

iexcl0

012

iexcl0

615

Aver

age

HR

090

5iexcl

058

8iexcl

059

6iexcl

061

71

032

iexcl0

270

044

91

88

6iexcl

179

7iexcl

163

51

83

6iexcl

201

81

310

005

1iexcl

038

2iexcl

110

6iexcl

031

11

283

Aver

age

BV

Pam

plit

ude

rela

tive

tolo

cal

base

line

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9iexcl

006

1iexcl

074

70

334

001

50

164

26

23

iexcl0

204

iexcl1

775

087

51

79

6iexcl

128

92

70

8

Aver

age

BV

Pam

plit

ude

rela

tive

togl

obal

base

line

141

7iexcl

272

62

32

41

417

iexcl2

211

28

86

iexcl2

529

iexcl1

367

159

01

110

iexcl1

165

23

01

19

17

iexcl2

664

18

84

20

82

iexcl1

759

27

58

Nu

mbe

ro

fte

mp

erat

ure

jum

ps

098

82

82

00

650

17

77

48

68

iexcl2

482

iexcl1

010

iexcl1

435

NA

NA

NA

164

40

837

29

02

114

52

44

52

63

3iexcl

321

3

Aver

age

resp

irat

ion

rate

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4iexcl

155

61

109

012

10

866

iexcl0

102

009

30

079

iexcl0

552

032

3iexcl

063

1iexcl

217

0iexcl

152

8

Aver

age

resp

irat

ion

ampl

itud

e

iexcl1

646

iexcl0

250

013

60

034

iexcl1

487

iexcl3

499

108

70

265

iexcl1

777

iexcl0

657

012

1iexcl

060

7iexcl

274

9iexcl

131

92

14

20

055

iexcl2

832

iexcl4

533

Bo

lded

stat

isti

csar

esi

gnifi

can

tat

the

5le

vel

Th

ele

ftp

anel

give

sag

greg

ate

resu

lts

for

allt

rad

ers

Sim

ilar

ttes

tsfo

rtr

ader

sw

ith

hig

hex

per

ienc

ean

dlo

wo

rm

oder

ate

exp

erie

nce

are

show

nin

the

mid

dle

and

righ

tp

anel

sre

spec

tive

lyE

ntr

ies

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

wer

ep

rese

nt

inei

ther

the

even

tor

cont

rol

feat

ure

vect

or

for

the

par

ticu

lar

mar

ket-

even

tty

pe

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spo

nse

sH

R=

hea

rtra

te

BV

P=

blo

od

volu

me

pul

se

328 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

deviations and trend-reversals both the number of SCRsand the number of temperature jumps reached statisti-cally significant levels for three out of the five event typesBVP amplitude-related features were highly significantfor volatility eventsmdashsignificance levels less than 1 wereobtained for both BVP amplitude features for all threetypes of volatility events Because the results were sosimilar across the three types of volatility events (max-imum volatility price volatility and return volatility) wereport the results only for maximum volatility in Table 2Such patterns of autonomic responses may indicate thepresence of transient emotional responses to deviationsand trend-reversal events that occur within 10-sec win-dows while volatility events defined in 5-min windowselicited a tonic response in BVP

To explore potential differences between experi-enced and less experienced traders a sample of 10traders was divided into two groups each consistingof five traders the first containing highly experiencedtraders and the second containing traders with low ormoderate experience where the levels of experiencewere determined through interviews with the tradersrsquosupervisors The results for each of the two groupsare reported in the two right subpanels of Table 2labelled lsquolsquoHigh Experiencersquorsquo and lsquolsquoLow or ModerateExperiencersquorsquo The high-experience traders generallyexhibited low responses for deviations and trend-re-versal events with only two types of market events(spread deviations and spread trend-reversals) elicitingsignificant mean autonomic responses (average HR)However even for high-experience traders volatilityevents induced statistically significant mean responsesfor both SCR and BVP amplitude Less experiencedtraders showed a much higher number of significantmean responses in the number of SCRs BVP ampli-tude and number of temperature increases for devia-tions and trend-reversal events Table 2 shows thatsessions with low- and moderate-experience tradersyielded 13 t statistics significant at the 5 level whilesessions with high-experience traders yielded only fivet statistics significant at the 5 level For both sets oftraders volatility events were associated with signifi-cant BVP amplitude The differences in responsepatterns for deviations and trend-reversals observedin this case indicate that less experienced traders maybe more sensitive to short-term changes in the marketvariables than their more experienced colleagues

To address the third objective 10-sec intervals imme-diately preceding and immediately following each devia-tion and trend-reversal event were compared to control(ie no-event) time intervals Two separate t tests wereconductedmdashone for pre- and another for post-eventtime intervalsmdashfor each of the three types of deviationsand two types of trend-reversals Because the objectivewas to detect differences in how pre- and post-eventfeature vectors differed from the control we used thesame control feature vectors for both sets of t tests In all

cases the t tests were designed to test the null hypoth-esis that both pre- and post-event feature vectors werestatistically indistinguishable from the control featurevector Surprisingly these two sets of t tests yielded verysimilar results reported in Table 3 implying that noneof the physiological variables were predictors of antici-patory emotional responses These findings may be atleast partly explained by how the events were defined Inparticular the definitions of deviation and trend-reversalevents are those instances where the time seriesachieved a prespecified deviation from the time-seriesmean and its moving average respectively In theabsence of large jumps the values of the time seriesnear an event defined in this way are likely to becomparable to the event itself Therefore the tradersmay be responding to market conditions occurringthroughout the pre- and post-event period not just atthe exact time of the event hence the similarity be-tween the two periods More complex definitions ofevents may allow us to discriminate between pre- andpost-event physiological responses and we are explor-ing several alternatives in ongoing research

Finally to address the fourth objective the followingfour groups of financial instruments were formed on thebasis of similarity in their statistical characteristics

Group 1 EUR JPY

Group 2 GBP CAD CHF AUD NZD (other majorcurrencies)

Group 3 MXP BRL ARS COP VEB CLP (LatinAmerican currencies)

Group 4 SPU EDU (derivatives)

Group 1 is by far the most actively traded set offinancial instruments and accounts for most of thetrading activities of our sample of 10 traders hence itis not surprising that the pattern of statistical signifi-cance for Group 1 in Table 4 is almost identical to thepattern for all traders in Table 2 (the left panel) SCRis significant for price and return deviations temper-ature jumps are significant for spread and price devia-tions and both measures of BVP are significant forvolatility events For Group 2 none of the eightphysiological variables were significant for price orspread deviations although temperature changes andrespiration rates were significant for return deviationsand both BVP measures were significant for pricetrend-reversals and volatility events The financial in-struments in Group 2 were not the primary focus ofany of the traders in our sample which may explainthe different patterns of significance as compared withGroup 1 Group 4 contains the fewest significantresponses and this is consistent with the fact thatthese securities were traded by two of the mostexperienced traders in our sample Derivative secur-ities are considerably more complex instruments thanthe spot currencies requiring more skill and trainingthan the other groups of securities However SCR was

Lo and Repin 329

Tab

le3

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gPr

e-an

dPo

st-e

vent

Tim

eIn

terv

als

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

Pre

-eve

nt

Inte

rva

lP

ost-

even

tIn

terv

al

Phy

sio

logy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mbe

rof

SCR

resp

ons

es3

75

61

543

25

40

21

18

24

01

057

92

87

20

604

20

88

16

64

063

60

579

Aver

age

SCR

amp

litu

de0

700

iexcl0

454

iexcl1

793

051

11

068

088

0iexcl

069

50

887

018

6iexcl

247

81

136

088

0

Aver

age

HR

096

8iexcl

020

5iexcl

055

2iexcl

066

3iexcl

038

6iexcl

027

00

905

iexcl0

588

iexcl0

596

iexcl0

617

103

2iexcl

027

0

Aver

age

BVP

amp

litud

ere

lativ

eto

loca

lba

selin

e0

043

iexcl0

895

006

00

814

iexcl0

045

36

19

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9

Aver

age

BVP

amp

litud

ere

lativ

eto

glob

alba

selin

e1

183

iexcl2

794

23

97

131

7iexcl

225

62

88

61

417

iexcl2

726

23

24

141

7iexcl

221

12

88

6

Nu

mbe

rof

tem

per

atur

eju

mp

s1

110

27

67

050

52

15

83

92

5iexcl

248

20

988

28

20

065

01

77

74

86

8iexcl

248

2

Aver

age

resp

irat

ion

rate

126

1iexcl

165

0iexcl

025

41

303

iexcl1

999

iexcl0

064

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4

Aver

age

resp

irat

ion

amp

litud

eiexcl

162

0iexcl

006

70

561

006

8iexcl

180

7iexcl

349

9iexcl

164

6iexcl

025

00

136

003

4iexcl

148

7iexcl

349

9

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

l

Th

ele

ftp

anel

cont

ain

st

stat

isti

csfo

rp

re-e

vent

feat

ure

vect

ors

and

the

righ

tp

anel

con

tain

st

stat

isti

csfo

rp

ost-

even

tfe

atu

reve

cto

rs

both

test

edag

ain

stth

esa

me

set

of

con

tro

ls

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spon

ses

HR

=h

eart

rate

B

VP

=bl

ood

volu

me

pu

lse

330 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

deviations and trend-reversals both the number of SCRsand the number of temperature jumps reached statisti-cally significant levels for three out of the five event typesBVP amplitude-related features were highly significantfor volatility eventsmdashsignificance levels less than 1 wereobtained for both BVP amplitude features for all threetypes of volatility events Because the results were sosimilar across the three types of volatility events (max-imum volatility price volatility and return volatility) wereport the results only for maximum volatility in Table 2Such patterns of autonomic responses may indicate thepresence of transient emotional responses to deviationsand trend-reversal events that occur within 10-sec win-dows while volatility events defined in 5-min windowselicited a tonic response in BVP

To explore potential differences between experi-enced and less experienced traders a sample of 10traders was divided into two groups each consistingof five traders the first containing highly experiencedtraders and the second containing traders with low ormoderate experience where the levels of experiencewere determined through interviews with the tradersrsquosupervisors The results for each of the two groupsare reported in the two right subpanels of Table 2labelled lsquolsquoHigh Experiencersquorsquo and lsquolsquoLow or ModerateExperiencersquorsquo The high-experience traders generallyexhibited low responses for deviations and trend-re-versal events with only two types of market events(spread deviations and spread trend-reversals) elicitingsignificant mean autonomic responses (average HR)However even for high-experience traders volatilityevents induced statistically significant mean responsesfor both SCR and BVP amplitude Less experiencedtraders showed a much higher number of significantmean responses in the number of SCRs BVP ampli-tude and number of temperature increases for devia-tions and trend-reversal events Table 2 shows thatsessions with low- and moderate-experience tradersyielded 13 t statistics significant at the 5 level whilesessions with high-experience traders yielded only fivet statistics significant at the 5 level For both sets oftraders volatility events were associated with signifi-cant BVP amplitude The differences in responsepatterns for deviations and trend-reversals observedin this case indicate that less experienced traders maybe more sensitive to short-term changes in the marketvariables than their more experienced colleagues

To address the third objective 10-sec intervals imme-diately preceding and immediately following each devia-tion and trend-reversal event were compared to control(ie no-event) time intervals Two separate t tests wereconductedmdashone for pre- and another for post-eventtime intervalsmdashfor each of the three types of deviationsand two types of trend-reversals Because the objectivewas to detect differences in how pre- and post-eventfeature vectors differed from the control we used thesame control feature vectors for both sets of t tests In all

cases the t tests were designed to test the null hypoth-esis that both pre- and post-event feature vectors werestatistically indistinguishable from the control featurevector Surprisingly these two sets of t tests yielded verysimilar results reported in Table 3 implying that noneof the physiological variables were predictors of antici-patory emotional responses These findings may be atleast partly explained by how the events were defined Inparticular the definitions of deviation and trend-reversalevents are those instances where the time seriesachieved a prespecified deviation from the time-seriesmean and its moving average respectively In theabsence of large jumps the values of the time seriesnear an event defined in this way are likely to becomparable to the event itself Therefore the tradersmay be responding to market conditions occurringthroughout the pre- and post-event period not just atthe exact time of the event hence the similarity be-tween the two periods More complex definitions ofevents may allow us to discriminate between pre- andpost-event physiological responses and we are explor-ing several alternatives in ongoing research

Finally to address the fourth objective the followingfour groups of financial instruments were formed on thebasis of similarity in their statistical characteristics

Group 1 EUR JPY

Group 2 GBP CAD CHF AUD NZD (other majorcurrencies)

Group 3 MXP BRL ARS COP VEB CLP (LatinAmerican currencies)

Group 4 SPU EDU (derivatives)

Group 1 is by far the most actively traded set offinancial instruments and accounts for most of thetrading activities of our sample of 10 traders hence itis not surprising that the pattern of statistical signifi-cance for Group 1 in Table 4 is almost identical to thepattern for all traders in Table 2 (the left panel) SCRis significant for price and return deviations temper-ature jumps are significant for spread and price devia-tions and both measures of BVP are significant forvolatility events For Group 2 none of the eightphysiological variables were significant for price orspread deviations although temperature changes andrespiration rates were significant for return deviationsand both BVP measures were significant for pricetrend-reversals and volatility events The financial in-struments in Group 2 were not the primary focus ofany of the traders in our sample which may explainthe different patterns of significance as compared withGroup 1 Group 4 contains the fewest significantresponses and this is consistent with the fact thatthese securities were traded by two of the mostexperienced traders in our sample Derivative secur-ities are considerably more complex instruments thanthe spot currencies requiring more skill and trainingthan the other groups of securities However SCR was

Lo and Repin 329

Tab

le3

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gPr

e-an

dPo

st-e

vent

Tim

eIn

terv

als

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

Pre

-eve

nt

Inte

rva

lP

ost-

even

tIn

terv

al

Phy

sio

logy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mbe

rof

SCR

resp

ons

es3

75

61

543

25

40

21

18

24

01

057

92

87

20

604

20

88

16

64

063

60

579

Aver

age

SCR

amp

litu

de0

700

iexcl0

454

iexcl1

793

051

11

068

088

0iexcl

069

50

887

018

6iexcl

247

81

136

088

0

Aver

age

HR

096

8iexcl

020

5iexcl

055

2iexcl

066

3iexcl

038

6iexcl

027

00

905

iexcl0

588

iexcl0

596

iexcl0

617

103

2iexcl

027

0

Aver

age

BVP

amp

litud

ere

lativ

eto

loca

lba

selin

e0

043

iexcl0

895

006

00

814

iexcl0

045

36

19

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9

Aver

age

BVP

amp

litud

ere

lativ

eto

glob

alba

selin

e1

183

iexcl2

794

23

97

131

7iexcl

225

62

88

61

417

iexcl2

726

23

24

141

7iexcl

221

12

88

6

Nu

mbe

rof

tem

per

atur

eju

mp

s1

110

27

67

050

52

15

83

92

5iexcl

248

20

988

28

20

065

01

77

74

86

8iexcl

248

2

Aver

age

resp

irat

ion

rate

126

1iexcl

165

0iexcl

025

41

303

iexcl1

999

iexcl0

064

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4

Aver

age

resp

irat

ion

amp

litud

eiexcl

162

0iexcl

006

70

561

006

8iexcl

180

7iexcl

349

9iexcl

164

6iexcl

025

00

136

003

4iexcl

148

7iexcl

349

9

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

l

Th

ele

ftp

anel

cont

ain

st

stat

isti

csfo

rp

re-e

vent

feat

ure

vect

ors

and

the

righ

tp

anel

con

tain

st

stat

isti

csfo

rp

ost-

even

tfe

atu

reve

cto

rs

both

test

edag

ain

stth

esa

me

set

of

con

tro

ls

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spon

ses

HR

=h

eart

rate

B

VP

=bl

ood

volu

me

pu

lse

330 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le3

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

othe

sis

Tes

tsof

the

Diff

eren

cein

Mea

nA

uton

omic

Resp

onse

sD

urin

gPr

e-an

dPo

st-e

vent

Tim

eIn

terv

als

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pone

nts

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chT

ype

ofM

arke

tEv

ent

(Col

umns

)

Ma

rket

Eve

nts

Pre

-eve

nt

Inte

rva

lP

ost-

even

tIn

terv

al

Phy

sio

logy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mbe

rof

SCR

resp

ons

es3

75

61

543

25

40

21

18

24

01

057

92

87

20

604

20

88

16

64

063

60

579

Aver

age

SCR

amp

litu

de0

700

iexcl0

454

iexcl1

793

051

11

068

088

0iexcl

069

50

887

018

6iexcl

247

81

136

088

0

Aver

age

HR

096

8iexcl

020

5iexcl

055

2iexcl

066

3iexcl

038

6iexcl

027

00

905

iexcl0

588

iexcl0

596

iexcl0

617

103

2iexcl

027

0

Aver

age

BVP

amp

litud

ere

lativ

eto

loca

lba

selin

e0

043

iexcl0

895

006

00

814

iexcl0

045

36

19

012

8iexcl

085

50

688

140

8iexcl

052

33

61

9

Aver

age

BVP

amp

litud

ere

lativ

eto

glob

alba

selin

e1

183

iexcl2

794

23

97

131

7iexcl

225

62

88

61

417

iexcl2

726

23

24

141

7iexcl

221

12

88

6

Nu

mbe

rof

tem

per

atur

eju

mp

s1

110

27

67

050

52

15

83

92

5iexcl

248

20

988

28

20

065

01

77

74

86

8iexcl

248

2

Aver

age

resp

irat

ion

rate

126

1iexcl

165

0iexcl

025

41

303

iexcl1

999

iexcl0

064

072

9iexcl

117

30

671

115

5iexcl

162

8iexcl

006

4

Aver

age

resp

irat

ion

amp

litud

eiexcl

162

0iexcl

006

70

561

006

8iexcl

180

7iexcl

349

9iexcl

164

6iexcl

025

00

136

003

4iexcl

148

7iexcl

349

9

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

l

Th

ele

ftp

anel

cont

ain

st

stat

isti

csfo

rp

re-e

vent

feat

ure

vect

ors

and

the

righ

tp

anel

con

tain

st

stat

isti

csfo

rp

ost-

even

tfe

atu

reve

cto

rs

both

test

edag

ain

stth

esa

me

set

of

con

tro

ls

DEV

=d

evia

tion

T

RV

=tr

end-

reve

rsal

V

OL

=vo

lati

lity

SCR

=sk

inco

ndu

ctan

cere

spon

ses

HR

=h

eart

rate

B

VP

=bl

ood

volu

me

pu

lse

330 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le4

tSt

atis

tics

for

Tw

o-Si

ded

Hyp

oth

esis

Tes

tsof

the

Diff

eren

cein

Mea

nAu

tono

mic

Resp

onse

sD

urin

gM

arke

tEv

ents

Ver

sus

No-

Even

tC

ontr

ols

for

Each

ofth

eEi

ght

Com

pon

ents

ofth

ePh

ysio

logy

Feat

ure

Vect

ors

(Row

s)fo

rEa

chTy

pe

ofM

arke

tEv

ent

(Col

umns

)G

roup

edIn

toFo

urD

istin

ctSe

tsof

Fina

ncia

lIn

stru

men

ts

Ma

rket

Eve

nts

Gro

up

1E

UR

an

dJP

YG

rou

p2

Oth

erM

ajo

rC

urr

enci

esa

Gro

up

3La

tin

Am

eric

an

Cu

rren

cies

bG

rou

p4

Der

iva

tive

sc

Phy

siol

ogy

Fea

ture

sD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

mD

EV

P

rice

DE

V

Spre

ad

DE

V

Ret

urn

TR

V

Pri

ceT

RV

Sp

rea

dV

OL

Ma

xim

um

DE

V

Pri

ceD

EV

Sp

rea

dD

EV

R

etu

rnT

RV

P

rice

TR

V

Spre

ad

VO

LM

axi

mu

m

Nu

mb

erof

SCR

resp

on

ses

16

53

iexcl1

902

19

64

iexcl0

554

iexcl2

286

iexcl0

940

033

20

112

054

4iexcl

179

5iexcl

241

7iexcl

110

82

49

82

18

8iexcl

039

72

99

5iexcl

040

50

898

26

82

109

41

66

91

594

iexcl0

239

22

36

Ave

rage

SCR

amp

litu

de

iexcl1

639

062

30

279

iexcl1

756

109

8iexcl

111

2iexcl

050

30

540

113

8iexcl

219

0iexcl

010

60

896

111

61

279

iexcl1

525

iexcl1

802

22

20

iexcl0

986

iexcl0

610

123

60

930

iexcl1

513

024

2iexcl

151

0

Ave

rage

HR

073

1iexcl

000

4iexcl

126

8iexcl

181

5iexcl

069

8iexcl

161

2iexcl

099

1iexcl

145

2iexcl

103

4iexcl

132

6iexcl

030

8iexcl

108

92

92

1iexcl

143

2iexcl

107

0iexcl

100

2iexcl

141

61

183

081

6iexcl

032

20

577

049

20

515

iexcl0

169

Ave

rage

BV

Pam

plit

ude

rela

tive

to

loca

lb

asel

ine

056

2iexcl

136

8iexcl

080

51

273

088

52

58

7iexcl

092

20

220

145

72

32

3iexcl

127

52

30

8iexcl

098

2iexcl

000

12

82

81

454

059

82

19

2iexcl

059

2iexcl

303

2iexcl

395

3iexcl

089

4iexcl

190

4iexcl

143

3

Ave

rage

BV

Pam

plit

ude

rela

tive

tolo

cal

bas

elin

e

059

7iexcl

301

01

71

40

375

iexcl0

252

25

92

052

11

448

097

23

08

7iexcl

149

92

63

3iexcl

192

5iexcl

084

91

482

025

9iexcl

035

1iexcl

012

9iexcl

066

0iexcl

150

1iexcl

327

4iexcl

162

6iexcl

200

4iexcl

241

1

Nu

mb

erof

tem

per

atu

reju

mp

s

068

84

14

1iexcl

144

42

28

30

295

iexcl2

043

107

3iexcl

155

72

40

41

248

37

26

iexcl1

564

iexcl1

603

iexcl1

275

iexcl1

922

iexcl0

512

iexcl1

128

iexcl1

540

iexcl1

367

iexcl1

349

iexcl1

362

iexcl1

367

iexcl1

053

NA

Ave

rage

resp

irat

ion

rate

052

5iexcl

174

1iexcl

079

70

163

iexcl1

884

iexcl0

004

075

6iexcl

056

02

02

3iexcl

109

5iexcl

178

4iexcl

011

60

037

044

6iexcl

112

40

643

iexcl1

392

012

5iexcl

194

6iexcl

029

10

361

041

6iexcl

110

5iexcl

172

5

Ave

rage

resp

irat

ion

amp

litu

de

iexcl2

120

iexcl0

692

034

5iexcl

050

9iexcl

093

5iexcl

207

9iexcl

074

60

560

047

00

905

089

3iexcl

383

90

280

iexcl0

207

142

6iexcl

036

0iexcl

053

4iexcl

043

70

920

iexcl0

955

iexcl1

915

125

71

593

iexcl0

039

Bo

lded

stat

isti

csar

esi

gnifi

cant

atth

e5

leve

la M

XP

BR

LA

RS

CO

PV

EB

CLP

bM

XP

BR

LA

RS

CO

PV

EB

C

LP

c ED

U

SPU

Ent

ries

mar

ked

lsquolsquoNA

rsquorsquoin

dic

ate

that

no

valid

data

was

pre

sen

tin

eith

erth

eev

ent

orco

ntro

lfe

atu

reve

cto

rfo

rth

ep

arti

cula

rm

arke

t-ev

ent

typ

e

DE

V=

dev

iati

on

TR

V=

tren

d-r

ever

sal

VO

L=

vola

tilit

ySC

R=

skin

con

duc

tan

cere

spo

nses

H

R=

hea

rtra

te

BV

P=

bloo

dvo

lum

ep

uls

e

Lo and Repin 331

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

significant for both price and return deviations inGroup 4 which was not the case for the sample ofhigh-experience traders in Table 2 SCR was alsosignificant for volatility events in Group 4 which isconsistent with the central role that volatility plays inthe pricing and hedging derivative securities (Black ampScholes 1973 Merton 1973) Volatility is a moresophisticated aspect of the trading milieu requiringa higher degree of training to fully appreciate itsimplications for market trends and profit-and-lossprobabilities This may explain the fact that SCR issignificant for volatility events only among the high-experience traders in Table 2 and the derivativestraders in Table 4

DISCUSSION

Our findings suggest that emotional responses are asignificant factor in the real-time processing of financialrisks Contrary to the common belief that emotions haveno place in rational financial decision-making processesphysiological variables associated with the ANS exhibitsignificant changes during market events even for highlyexperienced professional traders Moreover the re-sponse patterns among variables and events differed inimportant ways for less experienced tradersmdashmeanautonomic responses were significantly highermdashsug-gesting the possibility of relating trading skills to certainphysiological characteristics that can be measured

More generally our experiments demonstrate thefeasibility of relating real-time quantitative changes incognitive inputs (financial information) to correspond-ing quantitative changes in physiological responses in acomplex field environment Despite the challenges ofsuch measurements a wealth of information can beobtained regarding high-pressure decision makingunder uncertainty Financial traders operate in a con-trolled environment where the inputs and outputs ofthe decisions are carefully recorded and where thesubjects are highly trained and provided with greateconomic incentives to make rational trading decisionsTherefore in vivo experiments in the securities tradingcontext are likely to become an important part of theempirical analysis of individual risk preferences anddecision-making processes In particular there is con-siderable anecdotal evidence that subjects involved inprofessional trading activities perform very differentlydepending on whether real financial capital is at risk or ifthey are trading with lsquolsquoplayrsquorsquo money Such distinctionshave been documented in other contexts eg it hasbeen shown that different brain regions are activatedduring a subjectrsquos naturally occurring smile and a forcedsmile (Damasio 1994) For this reason measuring sub-jects while they are making decisions in their naturalenvironment is essential for any truly unbiased study offinancial decision-making processes

In capturing relations between cognitive inputs andaffective reactions that are often subconscious and ofwhich subjects are not fully aware our findings may beviewed more broadly as a study of cognitive- emotionalinteractions and the genesis of lsquolsquointuitionrsquorsquo Decisionprocesses based on intuition are characterized by lowlevels of cognitive control low conscious awarenessrapid processing rates and a lack of clear organizingprinciples When intuitive judgments are formed largenumbers of cues are processed simultaneously and thetask is not decomposed into subtasks (HammondHamm Grassia amp Pearson 1987) Expertsrsquo judgmentsare often based on intuition not on explicit analyticalprocessing making it almost impossible to fully explainor replicate the process of how that judgment has beenformed This is particularly germane to financial trad-ersmdashas a group they are unusually heterogeneouswith respect to educational background and formalanalytical skills yet the most successful traders seemto trade based on their intuition about price swingsand market dynamics often without the ability (or theneed) to articulate a precise quantitative algorithm formaking these complex decisions (Niederhoffer 1997Schwager 1989 1992) Their intuitive trading lsquolsquorulesrsquorsquoare based on the associations and relations betweenvarious information tokens that are formed on a sub-conscious level and our findings and those in theextant cognitive sciences literature suggest that deci-sions based on the intuitive judgments require not onlycognitive but also emotional mechanisms A naturalconjecture is that such emotional mechanisms are atleast partly responsible for the ability to form intuitivejudgments and for those judgments to be incorporatedinto a rational decision-making process

Our results may surprise some financial economistsbecause of the apparent inconsistency with marketrationality but a more sophisticated view of the role ofemotion in human cognition (Rolls 1990 1994 1999)can reconcile any contradiction in a complete andintellectually satisfying manner Emotion is the basisfor a reward-and-punishment system that facilitates theselection of advantageous behavioral actions providingthe numeraire for animals to engage in a lsquolsquocost- benefitanalysisrsquorsquo of the various actions open to them (Rolls1999 Chap 103) From an evolutionary perspectiveemotion is a powerful adaptation that dramatically im-proves the efficiency with which animals learn from theirenvironment and their past

These evolutionary underpinnings are more thansimple speculation in the context of financial tradersThe extraordinary degree of competitiveness of globalfinancial markets and the outsize rewards that accrue tothe lsquolsquofittestrsquorsquo traders suggest that Darwinian selectionmdashfinancial selection to be specificmdashis at work in deter-mining the typical profile of the successful trader Afterall unsuccessful traders are generally lsquolsquoeliminatedrsquorsquo fromthe population after suffering a certain level of losses

332 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Our results indicate that emotion is a significant deter-minant of the evolutionary fitness of financial tradersWe hope to investigate this conjecture more formally inthe future in several ways a comparison between tradersand a control group of subjects without trading experi-ence or with unsuccessful trading experiences an anal-ysis of the psychophysiological responses of traders in alaboratory environment a more direct analysis of thecorrelation between autonomic responses and tradingperformance a more fundamental analysis of the neuralbasis of emotion in traders aimed at the function of theamygdala and the orbitofrontal cortex (Rolls 1992 1999Chap 44- 45) and a direct mapping of the neuralcenters for trading activity through functional magneticresonance imaging (Breiter et al 2001)

There are a number of caveats that must be kept inmind in interpreting our findings First because of thesmall sample size of 10 subjects our results are at bestsuggestive and promising not conclusive A more com-prehensive study with a larger and more diverse sampleof traders a more refined set of market events and abroader set of controls are necessary before coming toany firm conclusions regarding the precise mechanismsof the psychophysiology of financial risk processing andthe role of emotion in market efficiency

Second while we have documented the statisticalsignificance of autonomic responses during marketevents relative to control baselines we have not estab-lished specific causal factors for such responses Manycognitive tasks generate autonomic effects hence amarked change in any one psychophysiological indexmay be due to changes in cognitive processing forexample complex numerical computations rather thanmore fundamental affective responses For examplemore experienced traders in our sample may have dis-played lower autonomic response levels because theyrequired less cognitive effort to perform the same func-tions as less experienced traders which may have nothingto do with emotion Without a more targeted set of tasksor additional data on the characteristics of the subjects itis difficult to distinguish between the two One way toaddress this issue is to correlate a subjectrsquos autonomicresponses with his or her trading performance so as todevelop a more complete profile of the relation betweenpsychophysiological indices and the dynamics of thetrading process However because trading performanceis generally summarized by multiple characteristicsmdashtotal profit-and-loss volatility maximum drawdownSharpe ratio and so onmdashestimating a relation betweenautonomic responses and such characteristics may re-quire significantly larger sample sizes and longer obser-vation windows due to the lsquolsquocurse of dimensionalityrsquorsquo Itmay be possible to overcome this challenge to somedegree through a more carefully crafted experimentaldesign in which specific emotional responses are pairedwith specific trading performance measures namelyanxiety levels during consecutive sequences of losing

trades With a larger sample size we can also begin totrack groups of traders over a period of time and throughvarying market conditions By measuring their autonomicresponse profiles at different stages of their careers itmay be possible to determine more directly the role ofemotion in financial risk processing

Finally a much larger set of controls should accompanya larger sample of traders These controls should includepsychophysiological measurements for professional trad-ers taken outside their natural trading environmentmdashideally in a laboratory setting that is identical for alltradersmdashas well as corresponding measurements forsubjects with little or no trading experience in bothtrading and nontrading environments Along with psy-chophysiological data more traditional personality in-ventory data for example MMPI or Myers- Briggs mayprovide additional insights into the psychology of trading

We hope to address each of these issues in ourongoing research program to better understand thecognitive processes underlying financial risk processing

METHODS

Subjects

The 10 subjectsrsquo descriptive characteristics are summar-ized in Table 5 Based on discussions with their super-visors the traders were categorized into three levels ofexperience low moderate and high Five traders spe-cialized in handling client order flow (Retail) threespecialized in trading foreign exchange (FX) and twospecialized in interest-rate derivative securities (Deriva-tives) The durations of the sessions ranged from 49 to83 min and all sessions were held during live tradinghours typically between 8 am and 5 pm easterndaylight time

Physiological Data Collection

A ProComp+ data-acquisition unit and Biograph (Ver-sion 12) biofeedback software from Thought Technol-ogy were used to measure physiological data for allsubjects All six sensors were connected to a smallcontrol unit with a battery power supply which wasplaced on each subjectrsquos belt and from which a fiber-optic connection led to a laptop computer equippedwith real-time data acquisition software (see Figure 2)Each sensor was equipped with a built-in notch filter at60 Hz for automatic elimination of external power linenoise and standard AgCl triode and single electrodeswere used for SCR and EMG sensors respectively Thesampling rate for all data collection was fixed at 32 HzAll physiological data except for respiration and facialEMG were collected from each subjectrsquos nondominantarm SCR electrodes were placed on the palmar sites theBVP photoplesymographic sensor was placed on theinside of the ring or middle finger the arm EMG triodeelectrode was placed on the inside surface of the

Lo and Repin 333

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

forearm over the flexor digitorum muscle group andthe temperature sensor was inserted between the elasticband placed around the wrist and the skin surface Thefacial EMG electrode was placed on a masseter musclewhich controls jaw movement and is active duringspeech or any other activity involving the jaw Therespiration signal was measured by chest expansionusing a sensor attached to an elastic band placed aroundthe subjectrsquos chest An example of the real-time physio-logical data collected over a 2-min interval for onesubject is given in Figure 3

The entire procedure of outfitting each subject withsensors and connecting the sensors to the laptop re-quired approximately 5 min and was often performedeither before the trading day began or during relativelycalm trading periods Subjects indicated that the pres-

ence of the sensors wires and a control unit did notcompromise or influence their trading in any significantmanner and that their workflow was not impaired in anyway This was verified not only by the subjects but alsoby their supervisors Given the magnitudes of the finan-cial transactions that were being processed and theeconomic and legal responsibilities that the subjectsand their supervisors bore even the slightest interfer-ence with the subjectsrsquo workflow or performance stand-ards would have caused the supervisors or the subjectsto terminate the sessions immediately None of thesessions were terminated prematurely

Physiological Data Feature Extraction

An initial smoothing of the raw EMG signals (sampled at32 Hz) was performed with a moving-average filter oforder 23 If the level of the filtered forearm EMG signalexceeded a threshold of 075 mV both SCR and BVPreadings at this time were discarded because of the highprobability of artifacts for example typing graspingtelephone handsets or inadvertent physical disturban-ces to the sensors Similarly if the level of the filteredfacial EMG signal exceeded 075 mV the respirationsignal at this time was excluded from further processingA very small spatial displacement of the sensor or theelectrode was able to produce a different kind ofartifactmdashan abrupt change in the signal of the orderof 10- 20 standard deviations within 1

32 of a second Suchjumps did not have any physiological meaning hencethey were excluded from further analysis via adaptivethresholding Specifically our adaptive thresholdingprocedure involved marking all observations that

Table 5 Summary Statistics for All Subjects Individual Traderrsquos Characteristics Specialty Type and Number of Market Time-Series Collected During the Session Session Duration and Absolute Time (Eastern Standard Time) of the Start of the Session

TraderID Gender Experience Specialty Market data Available

Nu m ber of MarketTim e-Series Recorded

Session Du ration(m in and sec)

SessionStart Tim e

B33 F high retail major currencies 2 4903000 1320

B34 M high FX major currencies 3 8303200 0856

B35 M high retail major currencies 4 6603000 1102

B36 M high derivatives SampP500 futureseurodollar futures

3 7901600 1255

B37 M moderate FX Latin Americanmajor currencies

9 7000600 0917

B38 M low FX Latin Americanmajor currencies

3 7201200 1102

B39 M high derivatives SampP 500 futureseurodollar futures

9 6200300 0819

B310 M low retail major currencies 7 6000800 0947

B311 M moderate retail major currencies 7 5402500 1132

B312 M low retail major currencies 7 5900000 0910

ProC om p+

Facial EMG Triode electrode measures

activity of the masseter muscle (detects speech)

Respiration sensor Elastic strap measures

chest expansion

Forearm EMG Triode electrode measures activity of the flexor digitorum muscle group (detects finger and arm movement)

Temperature sensor - thermistor

SCR sensor and electrodes

BVP photoplesymographicsensor

Control unit Fiber optic cable to a

data acquisition laptop

Figure 2 Placement of sensors for measuring physiologicalresponses

334 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

differed by more than 10 standard deviations from alocal average (with both standard deviation and localaverage computed over the most recent 30 sec of datawhich at a sampling rate of 32 Hz yields 960 observa-tions) and replacing these outliers with the immediatelypreceding values This procedure was then repeateduntil all such artifacts were eliminated Finally irrelevanthigh-frequency signal components and noise were elim-inated through a low-pass filter that was individuallydesigned for each of the physiological variables Therelatively smooth nature of the SCR signals permittedthe elimination of all harmonics above 15 Hz while theperiodic structure of the BVP signals pushed the cut-offfrequency to 45 Hz Table 6 reports the means andstandard deviations of the signals measured by the six

sensors for each of the 10 subjects After preprocessingfeature vectors were constructed from SCR BVP tem-perature and respiration signals

Financial Data Collection

At the start of each session a common time-marker wasset in the biofeedback unit and in the subjectrsquos tradingconsole (a networked PC or workstation with real-timedatafeeds such as Bloomberg and Reuters) and softwareinstalled on the trading console (MarketSheet by Tibco)stored all market data for the key financial instrumentsin an Excel spreadsheet time-stamped to the nearestsecond The initial time-markers and time-stampedspreadsheets allowed us to align the market and

Figure 3 Typical physiologicalresponse data recorded by sixsensors for a single subject overa 2-min time interval

85

0 05 1 15 2

25575

242628

7980582

04 05 06

12 13

35753653725

15345

Time min

Time min

Time min

See 3ASee 3ASee 3ASee 3ASee 3A

See 3B

Skin Conductance Response

Arm EMG

Facial EM

Respiratio

Temperature

m m

m V

yen F

Table 6 Summary Statistics of Physiological Response Data for All Traders Means and SD for Each of the Six Sensors

Facial EMG ( m V) Forearm EMG ( m V) SCR ( m m ) TMP ( 8C) BVP () RSP ()

Trader ID Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD

B32 123 146 884 2569 332 138 307 013 2334 631 3420 169

B34 136 315 088 186 1154 455 314 039 2493 405 3328 115

B35 126 165 331 357 216 271 313 074 2500 660 3557 165

B36 250 406 104 213 1009 216 297 036 2487 380 3345 157

B37 273 1853 366 1884 396 173 287 067 2506 327 3104 243

B38 835 2012 151 193 1224 221 284 025 2509 761 2975 061

B39 196 264 036 185 2509 340 296 015 2510 672 3665 108

B310 126 073 175 275 199 047 322 029 2487 512 3693 153

B311 137 859 185 1022 2887 210 333 055 2521 882 3312 161

B312 164 119 108 822 359 071 292 008 2510 512 3132 079

EMG = electromyographic data SCR = skin conductance responses TMP = body temperature BVP = blood volume pulse RSP = respiration rate

Lo and Repin 335

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

physiological data to within 05 sec of accuracy Figure 4displays an example of the real-time financial datamdashtheeuroUS dollar exchange ratemdashcollected over a 60-mininterval

Financial Data Feature Extraction

Deviations of a time series Zt were defined as thoseobservations that deviated from the series mean by acertain threshold where the threshold was defined asa multiple k of the standard deviation s z of the timeseries Positive deviations were defined as those ob-servations Zt such that Zt gt Z + k s z and negativedeviations were defined as those observations Zt suchthat Zt lt Z iexcl k s z The value of the multiplier k variedwith the particular series and session and it wascalibrated to yield approximately 5- 10 events persession Because the volatility of financial time seriescan vary across instruments and over time a singlevalue of k for all subjects and instruments is clearlyinappropriate However the sole objective in ourcalibration procedure was to maintain an approxi-mately equal number of events for each time seriesin each session Deviation events were defined forprices spreads and returns Table 7 reports theaverage values of k for each of the 15 financialinstruments in our study which range from 010 forARS and BRL (the Argentinian peso and Brazilian real)to 203 for EUR (the euro) for price deviations 010for ARS BRL and MXP (the Argentine peso Brazilianreal and Mexican peso) to 285 for GBP (the Britishpound) for spread deviations and 001 for ARS (theArgentine peso) to 1500 for COP and MXP (theColombian and Mexican peso) for return deviations

Trend-reversal events were defined as instanceswhen a time series Zt intersected its 5-min moving-average MA5 min(Zt) (see eg Figure 4 Panel 4A)Positive and negative trend-reversals were defined asthose observations Zt such that Zt gt (1 + d )MA5 min(Zt)and Zt lt (1 iexcl d )MA5 min(Zt) respectively The param-

eter d also varied with the particular series and session(see Table 7) ranging from an average of 00001 to 01for price series and 0005 to 1575 for spreads Due tothe high-frequency sampling rate (1 sec) prices did notvary often from one observation to the next hencemost of the return values were zero making it difficultto define trends in returns Therefore we definedtrend-reversals only for prices and spreads excludingreturns from this event category

Three types of volatility events were defined for 5-mintime intervals indexed by j based on the followingstatistics

where Ptjdenotes the average price in the interval tj iexcl

300 to tj and Rt2 is the squared return between t iexcl 1

and t We refer to s j1 as lsquolsquomaximum volatilityrsquorsquo since it is

the difference between the maximum and minimumprices as a fraction of their average and s j

2 and s j3 are

the standard deviations of prices and returns respec-tively lsquolsquoPlusrsquorsquo and lsquolsquominusrsquorsquo volatility events were thendefined as those instances when

s lj gt hellip1 Dagger h dagger s l

jiexcl1 hellipPlus Eventdagger hellip4dagger

s lj lt hellip1 iexcl h dagger s l

jiexcl1 hellipMinus Eventdagger hellip5dagger

for l = 1 2 3 The parameter h was calibrated to yieldfive or less volatility events per session and ranged froman average of 010 to 200 (see Table 7) For volatilityevents we calibrated the threshold to yield five or fewerevents to ensure that the combined time intervalscontaining volatility events comprised less than 50 ofthe total session time

Test Statistics

For each of the eight types of events a sample mean ˆm j

was computed for that component Xjt of the featurevector [X1t X2t X8t] and compared to the sample meanmˆ j

c of the corresponding component of the control fea-ture vector [Y1t Y2t Y8t] according to the test statistic

z j ˆˆm j iexcl ˆm c

j2ˆs 2

j =n j

q hellip6dagger

where

ˆm j ˆ 1

n j

Xn j

tˆ1Xjt ˆm c

j ˆ 1

n j

Xn j

tˆ1Yjt hellip7dagger

0 10 20 30 40 50 60

10733

15 16 17 18 19 20

1072

1073

1074Ask

BidTime min

Time min

euro$

See 4A

Panel 4A

Figure 4 Typical real-time market data (euroUS dollar exchangerate) recorded during a 60-min time interval and the corresponding5-min moving-average

s 1j ˆ

maxtjiexcl300lt t micro tj Pt iexcl mintjiexcl300lt t micro tj Pt

12 hellipmaxtjiexcl300lt t micro tj Pt Dagger mintjiexcl300lt t micro tj Pt dagger

hellip1dagger

s 2j ˆ

1

300

X

tjiexcl300lt t microtj

hellipPt iexcl Ptjdagger2

shellip2dagger

s 3j ˆ

1

300

X

tjiexcl300lt t microtj

R2t

shellip3dagger

336 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

Tab

le7

Aver

age

Val

ues

ofth

ePa

ram

eter

sU

sed

toD

efin

eM

arke

tEv

ents

D

evia

tion

s( k

)T

rend

-Rev

ersa

ls(d

)an

dVo

latil

ity

Even

ts(h

)

Secu

rity

Iden

tifi

erM

ean

kfo

rP

rice

Mea

nk

for

Spre

ad

Mea

nk

for

Ret

urn

Mea

nd

for

Pri

ceM

ean

dfo

rSp

rea

dM

ean

dfo

rR

etu

rn

Mea

nh

for

Ma

xim

um

Vol

ati

lity

Mea

nh

for

Ave

rage

Pri

ceV

ola

tili

ty

Mea

nh

for

Ave

rage

Ret

urn

Vol

ati

lity

ARS

010

010

001

010

000

010

0-

010

010

010

AUD

138

147

101

40

0009

00

475

-0

580

430

58

BRL

010

010

050

000

010

000

5-

200

010

00

200

0

CA

D1

020

8410

83

000

058

025

0-

072

072

063

CH

F1

752

129

170

0005

30

610

-0

380

420

38

CLP

040

100

120

00

0002

00

200

-0

600

500

50

CO

P0

500

5015

00

000

070

020

0-

500

300

300

EDU

140

250

110

00

0015

51

575

-0

750

700

43

EUR

203

243

706

000

066

127

4-

045

045

036

GB

P1

542

859

240

0003

90

472

-0

460

510

42

JPY

118

186

856

000

089

080

3-

054

054

041

MX

P1

000

1015

00

000

040

001

0-

100

105

085

NZD

158

130

100

00

0008

50

288

-0

730

650

73

SPU

063

025

140

90

0000

40

002

-0

680

070

03

VEB

190

250

110

00

0006

00

500

-0

300

400

40

See

Equ

atio

ns1-

3in

the

text

Lo and Repin 337

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

ˆs 2j ˆ

12n j iexcl 2

Xn j

tˆ1

hellipXjt iexcl ˆm jdagger2 Dagger

Xn j

tˆ1

hellipYjt iexcl ˆm cj dagger

2

Aacute

hellip8dagger

Under the null hypothesis that Xjt and Yjt are inde-pendently and identically distributed normal randomvariables with mean m and variance s 2 the test statisticz j has a t distribution with 2n jiexcl2 degrees of freedom Inthe absence of normality (Riniolo amp Porges 2000) theCentral Limit Theorem implies that under certain regu-larity conditions z j is approximately standard normal inlarge samples (Bickel amp Doksum 1977 Chap 44B) inwhich case the 95 critical region is defined by thecomplement of the interval [iexcl196 196]

Acknowledgments

Research support from the MIT Laboratory for FinancialEngineering and the Boston University Department ofCognitive and Neural Systems is gratefully acknowledged Wethank Jerome Kagan Ken Kosik JC Mercier Brett Steenbarg-er Svetlana Sussman and two reviewers for helpful commentsand discussion

Reprint requests should be sent to Andrew W Lo MIT SloanSchool of Management 50 Memorial Drive E52-432 Cam-bridge MA 02142 USA or via e-mail alomitedu

REFERENCES

Barber B amp Odean T (2001) Boys will be boys Genderoverconfidence and common stock investment Qu arterlyJou rnal of Econom ics 116 261- 292

Bechara A Damasio H Tranel D amp Damasio A R (1997)Deciding advantageously before knowing the advantageousstrategy Science 275 1293- 1294

Bell D (1982) Risk premiums for decision regretManagem ent Science 29 1156- 1166

Bickel P amp Doksum K (1977) Mathem atical statistics Basicideas and selected topics San Francisco Holden-Day

Black F amp Scholes M (1973) Pricing of options and corpo-rate liabilities Jou rnal of Political Econom y 81 637- 654

Boucsein W (1992) Electroderm al activity New YorkPlenum

Breiter H Aharon I Kahneman D Dale A amp Shizgal P(2001) Functional imaging of neural responses toexpectancy and experience of monetary gains and lossesNeuron 30 619- 639

Brown J D amp Huffman W J (1972) Psychophysiologicalmeasures of drivers under the actual driving conditionsJou rnal of Safety Research 4 172- 178

Cacioppo J T Tassinary L G amp Bernt G (Eds) (2000)Handbook of psychophysiology Cambridge UK CambridgeUniversity Press

Cacioppo J T Tassinary L G amp Fridlund A J (1990) Theskeletomotor system In J T Cacioppo amp L G Tassinary(Eds) Principles of psychophysiology Physical socialand in feren tial elem ents (pp 325- 384) Cambridge UKCambridge University Press

Clarke R Krase S amp Statman M (1994) Tracking errorsregret and tactical asset allocation Jou rnal of PortfolioManagem ent 20 16- 24

Critchley H D Elliot R Mathias C J amp Dolan R (1999)

Central correlates of peripheral arousal during a gamblingtask Abstracts of the 21st In ternational Sum m er School ofBrain Research Amsterdam

Damasio A R (1994) Descartesrsquo error Em otion reason andthe hu m an brain New York Avon Books

DeBond W amp Thaler R (1986) Does the stock marketoverreact Jou rnal of Finance 40 793- 807

Ekman P (1982) Methods for measuring facial action InK Scherer amp P Ekman (Eds) Handbook of m ethods innon verbal behavior research Cambridge UK CambridgeUniversity Press

Elster J (1998) Emotions and economic theory Jou rnal ofEconom ic Literature 36 47- 74

Fischoff B amp Slovic P (1980) A little learning Confidencein multicue judgment tasks In R Nickerson (Ed) Attentionand perform ance VIII Hillsdale NJ Erlbaum

Frederikson M Furmark T Olsson M T Fischer HAndersson J amp Langstrom B (1998) Functional neuro-anatomical correlates of electrodermal activity A positronemission tomographic study Psychophysiology 35 179- 185

Gervais S amp Odean T (2001) Learning to be overconfidentReview of Financial Studies 14 1- 27

Grossberg S amp Gutowski W (1987) Neural dynamics ofdecision making under risk Affective balance and cognitive-emotional interactions Psychological Review 94 300- 318

Hammond K R Hamm R M Grassia J amp Pearson T(1987) Direct comparison of the efficacy of intuitive andanalytical cognition in expert judgment IEEE Transactionson System s Man and Cybernetics SMC-17 753- 770

Huberman G amp Regev T (2001) Contagious speculation anda cure for cancer A nonevent that made stock prices soarJou rnal of Finance 56 387- 396

Kahneman D amp Tversky A (1979) Prospect theory Ananalysis of decision making under risk Econom etrica 47263- 291

Lichtenstein S Fischoff B amp Phillips L D (1982)Calibration of probabilities The state of the art to 1980In D Kahneman P Slovic amp A Tversky (Eds) Judgm entunder uncertain ty Heuristics an d biases (pp 306- 334)Cambridge UK Cambridge University Press

Lindholm E amp Cheatham C M (1983) Autonomic activityand workload during learning of a simulated aircraft carrierlanding task Aviation Space and En vironm entalMedicine 54 435- 439

Lo A W (1999) The three Prsquos of total risk managementFinancial An alysts Jou rnal 55 12- 20

Loewenstein G (2000) Emotions in economic theory andeconomic behavior Am erican Econom ic Review 90426- 432

Lorig T S amp Schwartz G E (1990) The pulmonary systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tialelements (pp 580- 598) Cambridge UK CambridgeUniversity Press

McIntosh D N Zajonc R B Vig P S amp Emerick S W(1997) Facial movement breathing temperature and affectImplication of the vascular theory of emotional efferenceCogn ition and Em otion 11 171- 195

Merton R (1973) Theory of rational option pricing BellJou rnal of Econom ics and Managem ent Science 4141- 183

Niederhoffer V (1997) Education of a specu lator New YorkWiley

Odean T (1998) Are investors reluctant to realize their lossesJou rnal of Finance 53 1775- 1798

Papillo J F amp Shapiro D (1990) The cardiovascular systemIn J T Cacioppo amp L G Tassinary (Eds) Principles ofpsychophysiology Physical social and in feren tial

338 Jou rnal of Cogn itive Neuroscience Volum e 14 Nu m ber 3

elements (pp 456- 512) Cambridge UK CambridgeUniversity Press

Peters E amp Slovic P (2000) The springs of action Affectiveand analytical information processing in choice Personalityand Social Psychology Bu lletin 26 1465- 1475

Rimm-Kaufman S E amp Kagan J (1996) The psychologicalsignificance of changes in skin temperature Motivation andEm otion 20 63- 78

Riniolo T amp Porges S (2000) Evaluating group distributionalcharacteristics Why psychophysiologists should beinterested in qualitative departures from the normaldistribution Psychophysiology 37 21- 28

Rolls E (1990) A theory of emotion and its application tounderstanding the neural basis of emotion Cogn ition andEm otion 4 161- 190

Rolls E (1992) Neurophysiology and functions of the primateamygdala In J P Aggelton (Ed) The am ygdalaNeurobiological aspects of em otion m em ory and mentaldysfunction (pp 143- 166) New York Wiley-Liss

Rolls E (1994) A theory of emotion and consciousness and its

application to understanding the neural basis of emotionIn M S Gazzaniga (Ed) The cogn itive neurosciences(pp 1091- 1106) Cambridge MIT Press

Rolls E (1999) The brain and em otion Oxford UK OxfordUniversity Press

Samuelson P A S (1965) Proof that properly anticipatedprices fluctuate randomly Indu strial Managem ent Review6 41- 49

Schwager J D (1989) Market wizards In terviews with toptraders New York New York Institute of Finance

Schwager J D (1992) The new m arket wizardsConversations with Am ericarsquos top traders New YorkHarperBusiness

Shefrin H (2001) Behavioral finan ce Cheltenham UKEdward Elgar Publishing

Shefrin M amp Statman M (1985) The disposition to sellwinners too early and ride losers too long Theory andevidence Jou rnal of Finance 40 777- 790

Tversky A amp Kahneman D (1981) The framing of decisionsand the psychology of choice Science 211 453- 458

Lo and Repin 339