Burnside 2011

Embed Size (px)

Citation preview

  • 8/11/2019 Burnside 2011

    1/22

    American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to The American Economic

    Review.

    http://www.jstor.org

    merican Economic ssociation

    The Cross Section of Foreign Currency Risk Premia and Consumption Growth Risk: CommentAuthor(s): Craig BurnsideSource: The American Economic Review, Vol. 101, No. 7 (DECEMBER 2011), pp. 3456-3476Published by: American Economic AssociationStable URL: http://www.jstor.org/stable/41408746

    Accessed: 14-08-2014 21:01 UTC

    Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available athttp://www.jstor.org/page/info/about/policies/terms.jsp

    JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of contentin a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.For more information about JSTOR, please contact [email protected].

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/http://www.jstor.org/action/showPublisher?publisherCode=aeahttp://www.jstor.org/stable/41408746http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/stable/41408746http://www.jstor.org/action/showPublisher?publisherCode=aeahttp://www.jstor.org/
  • 8/11/2019 Burnside 2011

    2/22

    Americanconomic

    eview

    01December

    011):

    456-3476

    http:/Avww.aeaweb.org/articles.php?doi=10.1257/aer.

    01.

    .3456

    The Cross Section of Foreign Currency Risk Premia and

    Consumption

    Growth

    Risk: Comment*

    By

    Craig

    Burnside*

    Hanno

    Lustig

    and

    Adrien Verdelhan

    2007)

    claim that

    aggregate

    consumption

    growth

    isk

    xplains

    theexcess

    returns o

    borrowing

    S dollarsto finance

    ending

    n

    other urrencies.

    hey

    reach this conclusion after

    stimating

    consumption-based

    asset

    pricing

    model

    using

    data on

    the returns f

    portfolios

    f

    short-term

    oreign-

    currency

    enominated

    money

    market ecurities orted

    according

    to their nterest

    differential

    ith he United States. Based on their

    vidence and additionalUS

    data,

    I

    argue

    that

    onsumption

    isk

    explains

    none of the cross-sectional

    variation n the

    expected

    returns f their

    ortfolios.

    Standard

    theorypredicts

    that

    the

    expected

    excess return f an

    asset,

    E(Ret),

    s

    given

    by

    -co

    v(Re

    m,

    ,

    where

    m,

    denotes some

    proposed

    stochastic iscount

    factor

    (SDF).

    Therefore,

    ny

    risk-based

    xplanation

    f the cross-section

    f returns elies

    on

    significant

    pread,

    across

    portfolios,

    n the covariance between the returns nd

    the

    SDF.

    For

    the

    SDFs

    that

    Lustig

    and Verdelhan

    henceforth,

    V)

    calibrate

    and

    estimate n their 007

    article,

    t s

    impossible

    to

    reject

    hat here s no

    spread

    n

    these

    covariances. n fact, t s impossibletorejectthat hese covariances are all zero.

    LV's SDF is linear

    n

    a vectorof risk

    factors,

    o

    they mplement widely

    used

    two-passprocedure

    o estimate ts

    parameters.

    he first

    ass

    is a series of time eries

    regressions

    f each

    portfolio's

    xcess return n the risk factors.These

    regressions

    determine he factor

    etas,

    .

    When there re

    n

    portfolios

    nd risk

    factors,

    is an

    nxk matrix. n LV's case n

    =

    8 and

    =

    3. None of the ndividual lements f LV's

    estimate,

    ,

    is

    statistically

    ifferentrom ero. For each of thethree

    actors,

    we also

    cannot

    reject

    the

    hypothesis

    hat ll

    eight

    of the relevant lementsof

    are

    ointly

    zero. Confrontedwith his

    vidence, alone,

    it would be reasonable to conclude that

    LV's model does not

    explain currency ortfolios

    orted

    n

    interest ates.

    The statisticalnsignificancef the factor etas impliesthatLV's measure of the

    SDF is also uncorrelatedwith the

    excess returns hat

    they study.

    To demonstrate

    this,

    considerthree

    alibrations f the

    parameters

    f the SDF in order o construct

    time series for

    mt:

    (i)

    the

    SDF

    parameters

    orresponding

    o LV's

    two-pass

    esti-

    mates,

    (ii)

    LV's

    Generalized Method of Moments

    (GMM,

    Lars

    P.

    Hansen

    1982)

    estimates f

    the SDF

    parameters,

    nd

    (iii)

    Motohiro

    Yogo's

    (2006)

    estimates f the

    *Duke

    niversity,

    epartment

    f

    conomics,urham,

    C

    7708,

    niversity

    f

    Glasgow,

    nd ationalureau

    of

    Economic

    esearch

    e-mail:

    [email protected]).

    thankohn

    ochrane,

    artin

    ichenbaum,

    avi

    Jagannathan,

    ergio

    ebelo,

    ichael

    eber,

    nd n

    nonymous

    eferee

    or

    elpful

    omments,

    nd he

    ational

    Science

    oundationor inancial

    upport

    SES-05

    6697).

    he sual

    isclaimer

    pplies.

    Toviewdditionalaterials,isithe rticleaget

    http://www.aeaweb.org/articles.php?doi=

    0. 57/aer.017.3456.

    3456

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    3/22

    VOL. 01

    NO.

    BURNSIDE:ISK REMIAND ONSUMPTION:OMMENT

    3457

    SDF

    parameters

    ased on stockreturns. then un series of time eries

    regressions

    of each

    portfolio's

    xcess return n the

    resulting

    m,

    series. n each

    case,

    I

    find hat

    estimatedSDF betas are

    jointly

    statistically

    ero.

    This,

    again, suggests

    that LV's

    model does notexplainthe returns o their urrency ortfolios.

    The second

    pass component

    of LV's estimation

    procedure

    s a cross-sectional

    regression

    f

    average portfolio

    eturns n the betas. This

    regression

    etermines he

    lambdas, X,

    x 1

    vectorof factor isk

    premia.

    There are two

    problems

    withLV's

    estimates f X.

    First,

    hey

    ocus almost

    entirely

    n standard rrors orX that reat he

    betas,

    3,

    as known

    regressors,

    ather han

    generated egressors.

    With hese standard

    errors

    X

    appears

    to

    be

    statistically ignificant,

    o LV draw favorable nference

    bout

    theirmodel.1 But

    treating

    as

    known

    eads

    to a

    misleading

    evel of confidence

    n

    themodel. With

    onventionally

    alculated standard

    rrors

    Shanken,GMM)

    none

    of

    the estimated actor isk

    premia

    n LV's benchmarkmodel and none of the

    param-

    etersof thecorrespondingstimated DF is statisticallyignificant,xcept n cases

    where the model has

    verypoor

    fit.

    Bootstrapped

    5

    percent

    onfidence

    egions

    for

    these

    parameters

    lways encompass

    zero.

    Consequently,

    draw unfavorable

    nfer-

    ence where

    they

    do

    not.

    Second,

    forX to be

    identified,

    musthave full olumn

    rank.Because most

    of the

    elements f

    3

    are

    statistically

    lose to

    zero,

    statistical ests ndicate hat

    herank f

    is

    very

    ow,

    perhaps

    s low as 0. The identification

    roblem

    aises two

    mportant

    ssues.

    First,

    nd most

    mportantly,

    t weakens

    nferencen the sense that

    ests f the

    pricing

    errors ased on the second

    pass regressions

    ave little

    power

    to

    rejectmisspecified

    models

    (Raymond

    Kan and Chu

    Zhang

    1999a;

    Burnside

    2010).

    Second,

    confidence

    regions or

    stimates f thefactor isk

    premia,X,generated sing symptotic

    tandard

    errors,

    ecome unreliable

    Kan

    and

    Zhang 1999b).

    Using

    methods hat

    re robust o

    weak

    dentification,

    show that

    V's data contain lmost

    no informationbout

    X. This

    reinforcesheunfavorable

    nference draw

    regarding

    heirmodel.

    In their

    eply,

    V

    defend

    heir

    indings

    n

    fourmain

    grounds.

    First,

    hey

    discard

    most

    of

    my

    comment

    s an obscure discussion

    of

    sampling

    uncertainty

    s

    opposed

    to

    point

    estimates.

    t is true that

    do not

    dispute

    their

    point

    estimates;

    this

    com-

    ment

    s not a trivial

    eport

    n errors

    n LV's code for

    ordinary

    east

    squares

    (OLS).

    Unfortunately,

    owever,

    ntil

    ataseis are

    nfinitely

    arge,

    nferencewill

    nvolveboth

    point

    estimates nd

    standard rrors.

    n their

    original

    article,

    LV

    clearly

    recognize

    the

    mportance

    f statistical

    ignificance

    or nference.

    hey repeatedly

    efer o the

    statistical ignificance ftheir stimates nd to theresultsof statistical ests.Once

    inference

    s conducted

    properly,

    owever,

    here s little

    upport

    or

    LV's model.

    Second,

    they appeal

    to

    a robustness

    heck,

    described

    n their

    rticle,

    n which

    additional test

    assets

    (six

    equity portfolios

    nd five

    bond

    portfolios)

    re

    included

    in

    the

    model estimation.

    he

    inclusionof these

    test

    ssets,

    however,

    has little

    ffect

    on

    my

    conclusions.

    One

    still cannot

    reject

    the

    null

    hypothesis

    hat he

    covariances

    between the excess

    returns

    f LV's

    currency ortfolios

    nd

    the SDF

    are all zero.

    Thus,

    regardless

    f the statistical

    ignificance

    f

    the

    parameters

    hatdetermine

    he

    1 herere evenablesf arameterstimatesn heriginalrticle.he tandardrrorsomputedn he irst

    six ables

    reat

    he etas

    s known.

    he tandard

    rrors

    omputed

    n heast able

    ffectively

    reathe etass

    unknown,

    ut

    hey

    annot

    e

    ompared

    o he

    tandard

    rrors

    n he

    estf

    he

    rticle,

    ecause

    hey

    re

    alculated

    for different

    odel

    one

    without

    constant).

    hey

    re

    lso alculated

    ncorrectly,

    s

    explain

    elow.

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    4/22

    3458 THE

    MERICANCONOMIC

    EVIEW DECEMBER011

    factor

    oadings

    n

    the

    SDF,

    I

    cannot

    reject

    hat hemodel

    predicts

    hat

    (Re,)

    =

    0 for

    all of the

    currency ortfolios. ncluding

    more

    testassets leads to modest

    mprove-

    menton the dentification

    ront,

    argely

    because the

    equityportfolios

    re

    correlated

    with one of the model's risk factors: he return o the aggregateUS stock market.

    Not

    surprisingly,

    his eads to some estimates f themodel

    parameters eing

    statisti-

    cally significant.

    owever,

    the model stilldoes not

    explain

    the

    cross-section f for-

    eign currency

    isk

    premia.

    The

    R2

    for he

    currency

    ortfolios

    lone is at best

    roughly

    zero,

    indicating

    hat he

    model

    cannot

    explain why

    some

    currency ortfolios

    ave

    significantlyositive

    returns

    while othershave

    significantly egative

    eturns.

    Third,

    LV refer o

    empirical

    evidence

    not

    in

    their

    original

    article. As in their

    more recent

    paper

    with Nick Roussanov

    (forthcoming),

    hey

    construct

    new set

    of seven

    portfolios

    rom heir

    riginal

    set of

    eightportfolios y

    considering

    trat-

    egies whereby

    the investor hort ells the low interest ate

    portfolio

    while

    going

    long in one of the seven higher nterest ateportfolios.Evidence regarding he

    seven "differenced"

    ortfolios

    ffectsnone of

    my

    conclusions. Most

    important,

    since these

    portfolios

    re linear combinationsof the

    original portfolios,

    ne can-

    not

    reject

    the null

    hypothesis

    hatthe covariances

    between the excess returns f

    the "differenced"

    ortfolios

    nd the SDF are all

    zero, and, therefore,

    ne cannot

    reject

    the null

    hypothesis

    hat he model

    predicts

    E(Ret)

    =

    0 for ll of the

    "differ-

    enced"

    portfolios.

    Also,

    because the

    "differenced"

    ortfolios

    re

    smaller

    n

    num-

    ber,

    nd are formed s linear

    combinations f the

    originalportfolios,

    working

    with

    these

    portfolios

    an

    only

    make the dentification

    roblem

    worse.

    Finally,

    LV

    bring

    o bear additional

    vidence based on the recent

    financial risis.

    They argue

    that he financial

    risis, lone,

    s sufficient

    vidence that heir

    onsump-

    tion-basedmodel works.

    ndeed,

    thefinancial risis

    s a

    single

    observation

    hat uits

    their

    hypothesis.

    Consumptiongrowth

    ell,

    and

    currency

    eturnswere

    negative,

    n

    late

    2008.

    However,

    show

    that

    carry

    rade

    returns nd

    consumptiongrowth

    re

    uncorrelated ver

    the full

    post-Bretton

    Woods

    period.

    Carry

    rade

    returns re cor-

    related with stock

    returns n the

    post-Bretton

    Woods

    period,

    but the

    market eta

    of the

    carry

    rade s far

    oo small to

    explain

    ts

    high average

    return. also

    establish

    that

    here s

    only

    a

    very

    weak

    tendency

    f the market

    eta of

    carry

    radereturns o

    increase

    during

    US

    recessions and

    periods

    of stock

    market urmoil. his

    casts doubt

    on

    any simple

    explanation

    f the

    returns o the

    carry

    radebased on

    market isk.

    I

    conclude

    that,

    aken

    s a

    whole,

    the

    vidence for

    LV's

    consumption-based

    model

    is extremelyweak. I cannotrejectthat hemodel-predictedxpectedreturns fthe

    currency ortfolios

    hey

    tudy

    re

    all zero. In their

    eply,

    V

    conclude

    by

    arguing

    that

    had the

    researchers f

    25

    years

    ago

    been

    confrontedwith their

    results,

    here

    would never

    have been a

    "forward

    remium uzzle."

    I

    am alive

    now,

    nd I have

    read

    their

    rticle.The

    forward

    remium

    s still

    puzzle.

    In

    Section

    I,

    I

    briefly

    eview LV's

    model,

    data,

    and

    methodological

    pproach.

    n

    Section

    I,

    I

    present

    he

    first-pass

    stimates f

    thebetas

    thatunderlie

    heir

    stimates

    of

    the factor

    isk

    premia

    and

    demonstrate

    hat here s

    little

    vidence of

    significant

    covariance

    between

    the

    portfolio

    eturns

    nd the risk

    factors. n

    Section

    III,

    I

    dis-

    cuss the

    second-pass

    estimates

    of the

    factorrisk

    premia

    and

    the

    interpretation

    f

    thepricing rrors nd calculate standard rrors orfactor iskpremiathat orrectly

    account for

    estimation

    f the

    betas. I

    discuss

    robustness f

    my

    negative

    findings

    n

    Section

    V. Section

    V

    concludes.

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    5/22

    VOL 01NO.

    BURNS1DE:

    ISK

    REMIAND

    ONSUMPTION:OMMENT

    3459

    I.

    Model, Data,

    Estimation,

    nd Inference

    LV work

    primarily

    with a

    log-linearized

    version of Motohiro

    Yogo's

    (2006)

    model, nwhich the stochastic iscountfactors given by

    (1)

    m,

    =

    [1

    -

    bc(Ac,

    -

    )

    -

    bd{Adt

    -

    )

    -

    br(rm

    -

    /ir)].

    Here

    c,

    represents

    he

    logarithm

    f a

    representative

    ousehold's

    consumption

    f

    nondurable

    goods,

    d,

    is the

    logarithm

    f

    the household's durable

    consumption,

    rWt

    s the

    logarithm

    of the

    gross aggregate

    return to

    wealth,

    c

    =

    E(Act),

    Hd

    =

    E(Adt),

    and

    =

    E(rWt).

    LV

    study

    the returns o

    borrowing

    US dollars

    in

    the

    money

    market o finance

    short-termecurities enominated n

    foreign urrency. hey

    form

    ightportfolios

    f

    suchpositions,which re createdby sortinghecurrencies ccording o their nterest

    differentialersusthe

    United States. refer o these

    portfolios

    s

    PI

    , P2,

    .

    P8 with

    the order

    running

    rom ow interest atecurrencies o

    high

    nterest atecurrencies.2

    LV estimate he model

    by exploiting

    he null

    hypothesis

    hatthe

    approximated

    stochastic discount factor

    SDF),

    m

    prices

    the n

    x

    1 vector of

    portfolio

    xcess

    returns,

    f.

    The

    pricing quation

    s

    (2) E(Retmt)

    =

    0.

    I rewrite

    1)

    generically

    s

    (3)

    =

    [1

    -

    (f,

    -

    |x)'b],

    where

    f,

    s

    x

    1 vectorof risk

    factors,

    p,

    =

    ^f,),

    b

    is

    x

    1 vectorof coeffi-

    cients,

    nd

    s a scalar

    representing

    he mean of the SDF.

    A.

    The

    Beta

    Representation

    nd Two-Pass

    Regressions

    It follows from

    3)

    and

    (2)

    that

    (4)

    (R0

    =

    cov(R

  • 8/11/2019 Burnside 2011

    6/22

    3460

    THE MERICANCONOMIC

    EVIEW DECEMBER

    011

    represents

    he /th ow n

    .

    LV

    estimate

    he

    ystem

    f

    equations represented

    by

    (5)

    using equation-by-equation

    LS. Given

    (4),

    the second

    pass

    is a cross-sec-

    tional

    regression

    f

    average portfolio

    eturns n

    the estimated etas:

    (6)

    '

    =

    (3-

    +

    a,

    i

    =

    ,...,n,

    where

    Re

    -

    }^(=1

    Reit, ,

    is the OLS estimate of

    ,

    obtained

    in the first

    tage,

    and

    a,

    is a

    pricing

    rror. et

    the

    OLS

    estimator f X be X

    =

    (' (3)_1

    '

    R',

    where

    Re

    is an

    n X 1 vectorformed rom he ndividualmean returns. he

    model's

    pre-

    dicted

    mean returns re

    X

    and the

    pricing

    errors re the

    residuals,

    6t

    =

    R*

    -

    X.

    The model's fit s assessed

    using

    the

    following

    tatistic:

    (i'

    U

    /?*

    =

    1

    -

    ('

    -

    X)'(R'

    -

    X)

    (i'

    U

    =

    1

    -

    (Re - Re)'(Re - Re)

    '

    where

    Ft

    =

    iE"=i

    is thecross-sectional

    verage

    of themean returnsn thedata.

    The model is testedon the basis of the estimated

    ricing

    rrors

    sing

    the statistic

    C&

    =

    Ta.'

    l,

    where

    &

    is a consistent stimator or he

    asymptotic

    ovariance

    matrix f

    VTol

    and the nverse s

    generalized.

    JohnH. Cochrane

    (2005)

    discusses

    how to form

    and shows

    that

    C

    -

    >

    Xn-k

    It is common to include a constant n the

    second-pass regression

    s follows:

    (8)

    R'

    =

    7

    +

    -X

    +

    ,

    i

    =

    1

    ...,n.

    The

    constant,

    ,

    is often

    nterpreted

    s the model's

    pricing

    error or

    the risk-free

    rate,

    but this error s shared

    by

    all assets. The

    statistical

    rgument

    or

    running

    he

    regression

    without he

    constant s thatwe know with

    ertainty

    hat heexcess return

    to a risk-free

    sset,

    or

    any

    other ero-beta

    sset,

    s zero. One

    argument

    or

    ncluding

    the constant s the notion hat he

    risk-free ate s

    imperfectly

    easured as the real

    return n T-bills.

    B. GMM Estimation

    Cochrane

    (2005)

    describes a GMM

    procedure

    hat

    produces

    the same

    point

    esti-

    mates as thetwo-pass regressionmethod butallows forheteroskedasticity-robust

    inference.When the

    constant s included n the

    model the moment

    estrictionsre

    (9)

    E{R%

    -

    a

    -

    if,)

    =0,

    i

    =

    l,...,

    .

    (10)

    E[(Rl

    -

    a

    -

    :fr)ft']

    0,

    i

    =

    1,...,.

    (11)

    E{R%-

    7

    -

    |X)

    =

    0,

    i

    =

    When the

    constant s excluded from

    he

    model,

    the ast set of

    moment

    estrictions

    is replacedby

    (12)

    tf(eft-iX)=0,

    i

    =

    l,...,

    .

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    7/22

    VOL

    01

    NO.

    BURNSIDE:ISK REMIAND

    ONSUMPTION:

    OMMENT

    3461

    In both

    cases,

    an

    identity

    matrix s used to

    weight

    he moment onditions.

    The model can also be estimated

    sing

    a GMM

    procedure

    hat

    reats he SDF as

    the

    primary bject

    of interest. his

    procedure,

    escribed n more detail n

    Cochrane

    (2005), estimates hemodel, 3), usingthe moment onditions:

    (13)

    {Rf[l

    -

    (f,

    -

    |x)'b]}

    =

    0

    (14)

    E(

    f,

    -

    ji)

    =

    0.

    The

    parameter

    s unidentifiednd is set

    equal

    to 1.

    The moment ondition

    13)

    can

    also be

    modified o allow for common

    pricing

    rror

    cross assets:

    (15)

    {ftf[l

    -

    (f,

    -

    ji)'b]

    -

    7}

    =

    0.

    As

    described

    n

    the online

    Appendix,

    the GMM

    procedure

    based on

    (14)

    and

    (15)

    can be set

    up

    so that t s

    numerically

    dentical

    o

    the

    two-passregression

    method n

    terms f

    pricing

    rrors.3

    II. First-Pass

    stimates

    f

    Betas

    Like

    LV,

    I

    compute first-pass

    stimatesof

    the betas

    by running

    he east

    squares

    regressions

    escribed

    by

    (5).

    I

    compute

    standard rrors

    sing

    standard

    ystem

    OLS

    formulas,

    s well as GMM-based

    procedures.

    also calculate 95

    percent

    onfidence

    regions using

    a

    bootstrapprocedure. Using any

    of the these

    procedures,

    none of

    the24 estimated etas is

    individually

    tatistically ignificant

    t the 5

    percent

    evel.4

    More

    important,

    hen there s

    spread

    n

    the

    expected

    returns cross

    portfolios,

    here

    should also be

    statistically ignificantpread

    n thebetas across

    portfolios.

    With

    his

    in

    mind,

    we can test whether

    or each

    factor

    the

    eight

    factorbetas are

    ointly sig-

    nificantly

    ifferent

    rom ero. As Table

    1(A)

    indicates,

    t conventional

    ignificance

    levels one cannot

    reject

    he

    hypotheses

    hat

    y

    =

    j

    V

    ,

    and

    y

    =

    0 V

    ,

    for ach fac-

    tor

    =

    1

    k. Since the contribution f factor

    to the vectorof

    model-predicted

    expected

    returnss A

    the atter

    hypothesis

    ests

    mply

    thatone

    cannot

    reject

    the

    null that each factor's risk

    premium

    contributes

    othing

    to the

    model-predicted

    expected

    returns.

    In their eply,LV mistakenly rguethat have looked onlyat individualbetas,

    when,

    n

    fact,

    n

    every

    version

    of

    my

    comment,

    ncluding

    this

    one,

    I

    have

    tested

    for

    spread

    in

    the

    betas.

    Second,

    they argue

    that should have looked

    at univari-

    ate betas rather han multivariate

    etas. This is

    puzzling, given

    that multivariate

    betas are what

    appears

    in the beta

    representation,4),

    and what enters

    nto the

    second-pass regression.

    Certainly,

    ne can define the matrix

    of univariate

    betas,

    "

    =

    cov(Rf,f()Df

    '

    where

    Df

    is a

    matrixwiththe variances

    of the factors

    n the

    3If

    n

    stimate

    f

    X s

    computed

    s X

    =

    Xy

    ,

    where

    yis

    he

    ample

    ovariance

    atrix

    f

    his

    stimate

    s

    identical

    o he

    wo-pass

    stimatef

    X.The

    quivalence

    f he MM

    nd

    wo-passrocedures

    sdemonstrated

    n

    the nlineppendix.Due o

    pace

    imitations

    report

    ullablesf etasn he nline

    ppendix.

    heGMM-basedtandardrrors

    I

    present

    re

    omputed

    sing

    variant

    f he

    ARHAC

    rocedure

    escribed

    y

    Wouter. en

    aan

    nd ndrew

    T.Levin

    2000).

    use

    VARHACtandard

    rrorso ake

    ntoccount

    ossible

    erial

    orrelation

    n

    GMM

    rrors.

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    8/22

    3462

    THE MERICANCONOMIC

    EVIEW

    DECEMBER

    011

    Table Factor

    etas ndCovariances:ests

    or

    pread

    nd gainstero

    (A)

    Betas

    (B)

    Covariances

    Ac Ad

    rw

    Ac Ad

    rw

    Standardrror

    ype

    Tests

    or o

    pread

    p- alues)

    System-OLS

    0.738

    0.563 0.273

    - -

    GMM-VARHAC 0.838

    0.596 0.437 0.611 0.901 0.405

    Jointestsersusero

    p- alues)

    System-OLS

    0.813

    0.623 0.365

    - -

    GMM-VARHAC 0.799 0.668

    0.511 0.447 0.682 0.510

    NotesAnnual

    ata,

    953-2002.

    n

    part

    A)

    the

    egressionquation

    s

    Reit a

    +

    f'ti eit,

    where

    %

    sthe xcesseturnf

    ortfolio

    at ime

    ,

    t

    (

    Act

    A

    dt

    W{)

    c sreal

    er

    ouse-

    hold

    onsumption

    nondurables

    nd

    ervices)

    rowth,

    d

    s

    real

    er

    ouseholdurableon-

    sumptionrowth,

    nd

    w

    s he alue

    eighted

    S

    tock

    arket

    eturn.he

    ortfolios

    re

    qually

    weightedroups

    f hort-term

    oreign-currency

    enominated

    oney

    arketecuritiesorted

    accordingo heirnterestifferentialithhe nitedtates,hereI and 8 re he ortfo-

    lios

    with,

    espectively,

    he mallestnd

    argest

    nterestifferentials.he able

    eports-values

    for estsf he

    ypotheses

    hat

    =

    j

    V and

    =

    0 Vi for ach actor

    .

    In

    part

    B)

    the

    covariances

    etween

    he

    xcesseturnsf he

    ortfolios

    nd he

    actors,

    ov(i?f,JJ),

    re stimated

    by

    GMM.

    he

    able

    eports

    -

    alues

    or ests

    f he

    ypotheses

    hat

    ov(/?f,

    )

    =

    c;

    V/

    nd

    co

    (/?f

    fj)

    =

    0 V

    forach

    actor

    .

    diagonal,

    and zeros

    off-diagonal.

    his leads to an alternative eta

    representation

    in which

    (Rf)

    =

    "X"

    and X"

    =

    D^b.

    It

    is

    perhaps

    more

    straightforward,

    ow-

    ever,

    to work with the SDF

    representation

    (Rf)

    =

    cov(Rf, f,)b.

    The columns of

    cov(R?,f)

    are

    proportional

    o the columns of

    ".

    We can

    directly

    stimate

    the

    elements of

    cov(Rf, f,)

    by

    GMM and then test whether

    ov(Rf,

    ;)

    ,

    V

    ,

    and

    cov(Rf

    fj,)

    =

    0

    Vi.

    As Table

    1(B)

    indicates,

    hese

    hypotheses

    annotbe

    rejected

    t

    conventional

    ignificance

    evels.

    It

    makes littledifferencewhich betas we

    use, however,

    ecause what matters n

    the end is how these

    betas are reflected

    n

    the covariance between

    the SDF and the

    portfolio

    eturns.

    sing

    (3), (2)

    can be rewrittens

    (16)

    E( Rf)

    =

    -cov(Rf,

    m,)/E(mt).

    With henormalization = 1,(3) impliesthat (mt) = 1,so we can rewrite16) as

    (17)

    E( Rf)

    =

    -

    cov(Rj, mt)

    =

    m,

    where

    m

    =

    -

    cov(Rf, m,)/

    l,, 'm

    =

    a2m

    nd

    a2m

    s the

    variance of

    mt.

    Given

    the

    definition f

    m

    (is, a.

    _

    m

  • 8/11/2019 Burnside 2011

    9/22

    VOL.

    01

    NO.

    BURNSIDE:

    ISK REMIAND

    ONSUMPTION:

    OMMENT

    3463

    Table

    SDF

    Betas: ests or preadnd gainst ero

    Model

    i)

    Model

    ii)

    Model

    iii)

    Standardrror

    ype

    Testsor o

    pread

    p- alues)

    System-OLS

    0.508

    0.477 0.575

    GMM-VARHAC

    0.688

    0.589

    0.602

    Jointestsersus

    ero

    p- alues)

    System-OLS

    0.469 0.443

    0.505

    GMM-VARHAC

    0.306 0.170

    0.329

    Notes: nnual

    ata,

    953-2002.

    he

    egressionquation

    s

    R%

    a

    mtim

    f

    lt,

    here

    mt

    1

    -

    (f,

    f)'b,

    sthe

    ample

    ean

    f

    ,

    nd he ectortakesn ne

    f he

    ollowing

    three

    alues:

    i)

    b

    =

    (-

    21.0129.9

    .46)' corresponds

    o V's

    wo-pass

    stimateith

    con-

    stant),ii)

    =

    (37.0

    4.7

    .65)'

    corresponds

    oLV'sGMM

    stimateith o

    onstant)

    nd

    (iii)

    o

    (6.74

    3.3

    .31)'

    the

    alibrated

    odel).

    ere

    eit

    s

    a

    portfolio

    eturn,

    nd

    ,

    s the

    vector

    f

    actors,

    escribed

    n

    able .The able

    eports-

    aluesorestsf he

    ypotheses

    hat

    im m

    nd

    im

    0 V.

    measure the SDF

    betas,

    we need data for

    m

    which can be constructed

    sing

    (3),

    and values for he

    elements f the vectorb. Here

    I

    use three

    ersionsof b taken

    directly

    rom V's article.

    In Table

    2(i)

    I

    use the b

    vector

    corresponding

    o LV's

    two-pass

    estimates of

    X:

    bc

    -

    -21

    ,bd

    =

    130,

    and

    br

    =

    4.5.

    Table

    2(ii)

    uses theb vector

    orresponding

    o

    LV's GMM estimates f b:

    bc

    =

    37,

    bd

    =

    75,

    and

    br

    =

    4.7. Table

    2(iii) repeats

    he

    exercise

    using

    the

    b vector

    orresponding

    o the

    calibratedmodel discussed

    in sec-

    tion E of LV's article:bc

    =

    6.7, bd

    =

    23,

    and

    br

    =

    0.31. As Table

    2

    indicates,

    n

    all

    of thesecases the

    null

    hypotheses

    hat

    im

    m

    or ll i and

    im

    =

    0

    for ll i cannot

    be

    rejected.

    n

    other

    words,

    here s no

    spread

    n the

    betas,

    and

    they

    re

    ointly

    ero.

    Tests based

    directly

    n

    cov(Rf,

    m,

    rather han

    m

    each the ame

    conclusionbut

    re

    not

    reported

    n

    the

    table.

    Given

    that the SDF betas

    are

    jointly

    statistically

    nsignificant,

    conclude

    that

    LV's

    model does not

    explain

    the cross-section

    f the

    expected

    returns f

    their

    ort-

    folios.

    In Section

    IV I show that his

    finding

    s robust o

    (i)

    estimates

    of the SDF

    based on an

    expanded

    set of

    test assets

    including equities

    and

    bonds,

    (ii)

    using

    the seven

    "differenced"

    ortfolios

    mphasized

    in

    their

    eply,

    nd

    (iii)

    a

    higher

    fre-

    quency,post-Bretton

    Woods,

    developed-country

    atabase that

    xtends

    through

    he

    recentfinancial risis.Thus,themessage of this comment s immune o thepoints

    emphasized

    by

    LV in their

    eply.

    III. Second

    Pass and

    GMM Estimates

    fthe

    Model

    The second

    pass

    and GMM

    estimates

    of the model

    provide

    us another

    pportu-

    nity

    o assess

    LV's

    proposed

    explanation

    of the cross-section

    f returns

    o

    foreign

    currency ortfolios.

    Of

    particular

    nterest re

    the

    point

    estimates

    f X and

    b,

    the

    R2

    measure

    of fit nd the

    testsof the

    pricing

    rrors.

    LV's

    second-pass

    regression,

    which ncludes

    the onstant

    ,

    is

    reproduced

    n Table

    3.Whenpresentingheir indingshey howOLS standard rrors,which ssumethat

    the

    first-pass

    etas

    are known.

    Given these

    standard

    rrors,

    hefactor

    isk

    premia

    for

    consumption

    nd

    durables

    are both

    positive

    nd

    highly

    tatistically

    ignificant.

    he

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    10/22

    3464

    THE MERICANCONOMIC

    EVIEW

    DECEMBER

    011

    Table Second-Pass

    egression

    ith Constant

    Factor

    rices

    X)

    Constant

    Nondurables

    Durables

    Market

    R2

    p-

    alue MAE

    () {Ad) (rw)

    -2.94

    29 4/70

    333 (

    0.44

    (0.86) (0.83)

    (0.97) (7.59)

    (0.483)

    [2.23]

    [2.11] [2.42]

    [18.8] [0.972]

    {2.66} {2.48}

    {2.41} {23.1}

    {0.994}

    Notes

    nnual

    data,

    1953-2002.

    he table

    eports

    esults rom

    unning

    he cross-sectional

    egres-

    sion

    *

    7

    +

    fX

    -

    where

    e

    sthemeanxcess

    eturnf

    ortfolio

    and

    ,

    sthe ectorf actoretas

    f

    portfolio

    estimated

    n he

    irst-pass

    egression.

    he

    ortfolios

    nd actorsre escribed

    nTable

    .

    For he ac-

    tor isk

    remia

    X)

    OLS

    tandardrrors

    re n

    parentheses,

    hanken

    tandardrrorsre

    n

    quare

    rackets,

    nd

    GMM-VARHAC

    tandardrrorsre

    n races.

    ootstrapped

    5

    percent

    onfidence

    egions

    re

    n

    ngled

    rackets.

    For he ests

    f he

    ricing

    rrors

    compute

    he esttatistic

    orach f he hree ethods

    f

    omputing

    he

    ovari-

    ancematrix

    f

    OLS, hanken,

    nd

    MM-VARHAC)

    nd

    eport

    he ssociated

    -value.

    he statisticromhe

    second-passegressionsreportedlong ithhemeanbsolutericingrrorMAE).

    R2 of the model is 0.87 and the

    p-value

    forthe

    test for

    significance

    f the

    pricing

    errors

    s

    0.48. These results

    re a

    key

    basis of LV's

    positive

    ssessment f themodel.

    There are three easons our assessment hould be less

    sanguine.

    The

    main one is

    thatOLS standard rrors re

    nappropriate iven

    heestimation f the

    betas,

    and this

    turns ut to matter

    great

    deal for nference.Once standard rrors re

    computed

    appropriately,

    stimatesof X and b are

    statisticallynsignificant.

    he latter

    inding

    is especially mportant ecause itsuggests hat heconsumption actors o nothelp

    price currency

    eturns. he second reason to be

    skeptical

    s that he model

    performs

    much more

    poorly

    when we

    impose

    therestrictionhat heconstant s

    equal

    to zero.

    The b

    parameters

    emain

    nsignificant,

    nd the fitof the model deteriorates ub-

    stantially.

    he third eason to be doubtful

    bout the model estimates s thatthere

    is a severe

    identification

    roblem.

    Under nonidentificationr weak

    identification,

    asymptotic

    tandard rrors

    even arguably ppropriate

    nes)

    are

    likely

    o

    understate

    the

    degree

    of

    uncertainty

    bout the model

    parameters.

    A.

    Inference

    bout Model Parameters

    As Cochrane (2005) pointsout,the fact that he betas are estimated n the first

    pass

    mattersfor

    inference bout the factor risk

    premia,

    and this remains true

    asymptotically.

    here are three tandard

    ways

    to

    deal withthis

    problem.

    One is

    to

    use the correction

    f the standard rrors

    uggestedby Jay

    Shanken

    1992).

    Another

    is to

    compute

    tandard rrors

    sing

    the

    first f the two GMM

    procedures

    described

    above,

    because it

    produces

    the

    same

    point

    estimates.A third

    s to construct on-

    fidence

    regions

    for the

    parameters sing

    bootstrap

    methods.

    By

    construction,

    he

    alternative

    pproaches

    o

    calculating

    tandard

    rrors o not ffect he

    point

    stimates

    of the

    factor isk

    premia.

    The three

    procedures

    ead to similar nference

    egarding

    the

    model,

    and

    using

    them,

    ather

    han OLS

    standard

    rrors,

    matters

    oth

    qualita-

    tively ndquantitatively.

    The Shanken

    and

    GMM-corrected standard

    rrorsfor the

    model with the con-

    stant

    Table

    3)

    are

    roughly

    wo

    to three imes

    arger

    han the

    OLS standard rrors

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    11/22

    VOL 01

    NO.

    BURNSIDE:ISK REMIAND

    ONSUMPTION:OMMENT

    3465

    that

    gnore

    estimationof the

    betas.

    Why

    is the Shanken

    correction o

    big?

    Let

    0

    =

    (7

    X')',

    E

    =

    E(ete't),

    and let

    f

    be a matrixwith

    a

    leading

    column and row

    of

    zeros,

    and

    Ef

    in the ower

    right

    orner.When the betas are

    treated s known he

    covariancematrix fV (0 - 0) is

    (19)

    e

    =

    (+'

    V

    +'

    E+(+' +1

    +

    f.

    Here

    +

    =

    (t )

    and l is an n x 1

    vector f ones. With he

    Shankencorrection he

    covariancematrix s

    (20)

    n

    =

    (1

    +

    X'

    Ef1

    X)(+'

    V

    0+'

    S+(+' +)"'

    +

    f.

    In

    some finance

    pplications

    the Shanken

    correction s small. For

    example,

    forthe

    CAPM estimated singthe annual returns f Fama and KennethR. French's 1993)

    25

    portfolios

    ortedon

    size and book-to-market alue over the

    period

    1953-2002,

    the Shanken-correction

    erm,

    1

    +

    A2/

    },

    is estimated o be 1.03. In LV's case the

    estimate f

    1

    -I-

    X'

    Ef1

    X is

    6.79.

    Although

    he ndividualAs

    in

    LV's model are of

    the same orderof

    magnitude

    s for he

    CAPM,

    the

    consumption

    actors ave much

    smallervariancethanthe market eturn.

    his blows

    up

    the size

    of

    the Shanken cor-

    rection

    ubstantially.

    Using

    either he Shanken or GMM standard

    rrors,

    none of the estimatedfac-

    tor risk

    premia

    in Table 3 are

    statistically ignificant

    t the 5

    percent

    evel. The

    bootstrap-based

    5

    percent

    onfidence

    egions

    for he

    parameters

    lso

    encompass

    0.

    These resultsdo not

    mply

    that he

    price

    of

    consumption

    isk s zero.

    Instead,they

    indicate that he

    oint

    behavior of the

    currency

    eturns nd

    consumption

    actors

    s

    uninformativebout the

    price

    of

    consumption

    isk.

    LV defend he tatistical

    ignificance

    f their

    indings

    n three

    rounds.

    irst,

    hey

    appeal

    to Ravi

    Jagannathan

    nd

    Zhenyu

    Wang

    1998)

    to defend heuse

    of OLS stan-

    dard errors ather

    han the

    Shanken

    correction. his is

    inappropriate.

    agannathan

    and

    Wang's point

    s thatunder

    heteroskedasticity,

    he Shanken

    correction s

    inap-

    propriate,

    nd thatmore

    general

    GMM errors re

    appropriate.

    hanken's

    proof

    hat

    corrected tandard rrors re

    necessarilybigger

    thanOLS standard

    rrors oes not

    work forGMM standard rrors.GMM

    errors ould be smaller

    than OLS standard

    errors,

    ut n LV's case

    they

    re not.

    Second,

    in theirfootnote

    11,

    they rgue

    that

    theOLS standard rrors re close inmagnitude o theGMM standard rrors. his is

    because

    they nappropriatelyompare

    GMM standard rrors

    or he model

    without

    a

    constant,

    o the OLS standard rrors

    or model with constant.5

    he

    appropri-

    ate

    GMM

    standard rrors re

    actually

    close in

    magnitude

    o the Shanken

    standard

    errors.

    hird,

    hey rgue

    that tandard

    rrors rom

    bootstrap rocedure

    re small

    enough

    to make the estimated

    isk

    premia

    significant.

    ather

    than focus on boot-

    strap

    tandard

    rrors,

    use the

    entire istribution

    f

    bootstrapped

    stimates o show

    that 5

    percent

    onfidence

    egions asily

    encompass

    0.

    It is

    especially

    important

    o know

    whether he

    consumption

    actors

    help

    to

    price

    the

    currency

    eturns. his

    requires

    us to focus

    on the

    parameter

    ector b.

    GMM

    5

    When

    V

    eport

    hanken

    nd

    tandard

    rrors

    orhemodel

    ithout

    he

    onstant,

    heyppear

    o

    se ncor-

    rect

    ormulas,

    sdetailed

    n he nline

    ppendix.

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    12/22

    3466 THE MERICANCONOMIC

    EVIEW

    DECEMBER011

    Table

    A

    GMM

    stimates

    f heModelwith he

    Constant

    Model

    arameters

    b)

    Stage

    Nondurables

    Durables Market

    R2

    p-value

    MAE

    (Ac) (Ad) (rw)

    First

    -21.0

    129^9

    46 087 0.44

    (87.7) (97.4)

    (4.83)

  • 8/11/2019 Burnside 2011

    13/22

  • 8/11/2019 Burnside 2011

    14/22

    3468

    THE MERICANCONOMIC

    EVIEW

    DECEMBER

    011

    Table

    Second-Pass

    egressionithoutConstant

    Factor

    rices

    X)

    Nondurables

    urables Market

    R2

    /?-

    alue

    MAE

    (Ac) (A ) (rw)

    0.59

    1.10

    11.7 0.34

    1.17

    (0.73)

    (1.02) (7.40)

    (0.001)

    1.01] [1.40]

    [10.1]

    [0.059]

    {1.17}

    {1.69}

    {10.6}

    {0.173}

  • 8/11/2019 Burnside 2011

    15/22

    VOL. 01

    NO.

    BURNSIDE:ISK REMIA

    ND

    ONSUMPTION:OMMENT

    3469

    Table

    GMM

    stimatesf

    heModel

    with o

    Constant

    Factor

    rices

    X)

    Stage

    Nondurables

    Durables

    Market

    (Ac) (Ad) (rw)

    First

    0.59

    1.10

    11.7

    (1.07)

    (1.64)

    (8.26)

    Second

    2.37

    3.48

    10.2

    (1.00)

    (1.13)

    (7.37)

  • 8/11/2019 Burnside 2011

    16/22

    3470 THE MERICANCONOMIC

    EVIEW DECEMBER011

    my

    conclusion,

    based on betas

    alone,

    that he model cannot

    explain

    differencesn

    expected

    returns cross

    currency ortfolios.

    Adding

    the six Fama-French

    quityportfolios

    o the set of test ssets

    slightly

    lle-

    viates the dentificationroblembecause equitieshave statistically ignificantetas

    with

    respect

    o the market eturn

    actor,

    w.

    However,

    he rank ests

    presented

    n

    the

    online

    Appendix

    ndicate that he

    matrix

    till

    appears

    to

    have

    reduced

    rank.The

    identification

    roblem

    does not

    go away

    withthe furtherdditionof the five

    Fama

    bond

    portfolios.

    Estimatesof the model without he constant

    sing

    the

    currency

    nd

    equityport-

    folios as test assets are

    presented

    n the online

    Appendix.

    As indicated

    there,

    with

    sufficientterations ver the

    weighting

    matrix,

    he factor isk

    premia

    for

    onsump-

    tion

    growth

    nd durables

    growth

    re

    statistically ignificant.

    owever,

    he fit f the

    model with

    respect

    o

    currency ortfolios

    s

    verypoor.

    When theR2 statistic s cal-

    culated ust for urrency ortfoliostrangesfrom .03 (at the firsttageofGMM)

    to -0.76

    (for

    iterated

    GMM).

    The mean absolute

    pricing

    errorfor the

    currency

    portfolios anges

    from1.44

    (at

    thefirst

    tage

    of

    GMM)

    to 1.88

    (for

    terated

    GMM).

    Why

    do I

    compute

    these statistics

    ust

    for

    currency ortfolios?

    First,

    the

    goal

    is to

    explain

    the cross-section f returns f

    currency ortfolios.

    he

    R2

    across all

    assets does not tell us whether he model

    explains why

    some

    currency ortfolios

    earn

    high

    returns nd others arn ow returns.

    nstead,

    he

    R2

    across

    ust

    the

    currency

    portfolios

    ells us whether he model

    explains why

    some

    currency

    ortfolios

    like

    P7)

    earn

    high

    returns,

    nd other

    urrency ortfolios

    like PI)

    earn ow

    returns,

    n

    average.

    Second,

    we are not

    trying

    o

    explain

    why

    the

    currency ortfolios

    ll have

    relatively

    ow

    returns

    the average

    excess return

    cross LV's

    eightcurrency ort-folios is 0.1

    percent)

    compared

    to the

    equityportfolios

    the

    average

    excess returns

    of the six

    Fama-French

    portfolios

    re all above 6

    percent,

    nd

    they

    re centered

    near 9

    percent).

    That s not

    a

    puzzle, given

    that

    urrency ortfolios

    re

    only weakly

    correlatedwithrisk

    factors hat

    price equity portfolios.

    he

    puzzle

    is

    why

    the ow

    interest ate

    currency ortfolio,

    I,

    has an

    average

    return f -2.3

    percent,

    nd

    why

    the

    high

    nterest ate

    currency ortfolios,

    7 and

    P8,

    have

    average

    returnsn

    excess

    of

    2

    percent.

    Adding

    the fiveFama

    bond

    portfolios

    oes not

    mprove

    he

    situation.At the

    first

    two

    stages

    of GMM the

    results re

    quite

    similarto those

    obtained

    using only

    the

    currency

    nd

    equity

    portfolios,

    lthough

    urther

    terations ver

    the

    weighting

    matrix

    eventually rive utconsumption rowthnd durablesgrowths significantiskfac-

    tors.Once

    again,

    thefit f the

    model with

    espect

    o

    currency ortfolios

    s

    verypoor.

    If

    theR2

    statistic s

    calculated

    ust

    for

    urrency ortfolios

    t

    ranges

    from

    .03

    (at

    the

    first

    tage

    of

    GMM)

    to

    -

    1.35

    (for

    terated

    GMM).

    The mean absolute

    pricing

    rror

    for he

    currency

    ortfolios

    anges

    from1.40

    (at

    thefirst

    tage

    of

    GMM)

    to 1.64

    (for

    iterated

    GMM).

    B.

    Differenced

    urrency

    ortfolios

    In

    their

    eply,

    V

    argue

    that

    hey

    an

    explain

    the

    excess returns o

    the

    strategy

    f

    holdingPi andshortingI, for = 2, ,8. Theyclaimthat heir rticlesreally bout

    these even

    "differenced"

    ortfolios,

    hich

    refer o as

    D2,

    D3,

    .

    D8,

    and not

    really

    about

    he

    original

    ight ortfolios.

    iven

    that he

    ntire rticle s

    about he

    P-portfolios

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    17/22

    VOL

    01

    NO.

    BURNSIDE:

    ISK

    REMIAND

    ONSUMPTION:

    OMMENT

    3471

    this comment s

    surprising,specially given

    the

    following

    tatement

    n the

    original

    article:

    consumption-based

    models can

    explain

    thecross-section f

    currency

    xcess

    returns

    f and

    only

    f

    high

    nterest

    ate

    currencies

    ypically epreciate

    when real US

    consumption rowths low,while ow interest ate urrenciesppreciate."Notice that

    these tatementsre not bout whether here s a differenceetween herates f return

    of the

    portfolios;

    t s a statementbout therates f return hemselves.

    LV

    also

    argue

    for

    working

    with he

    D-portfolios

    n the basis that

    large swings

    n

    thedollarmake t hard o

    accurately

    stimate he

    onstant,"

    he onstant

    eing

    7

    in

    the

    model for he

    P-portfolios.

    his

    argument

    s not

    persuasive

    because the

    ntercept

    an

    be "estimated"with

    perfect ccuracy.

    We know hat he mean excess return f a zero

    beta asset s

    zero,

    so we can set

    7

    equal

    to zero without ven

    having

    o estimate t.

    Nonetheless,

    whatof LV's

    point

    hat he onstant o

    longerplays

    an

    important

    ole

    in

    explaining

    hecross-section nce we consider

    he even D

    portfolios?

    V are

    right,

    butthispoint an easilybe made without ewtables ofpoint stimates. onsider he

    model with heconstant. he estimates

    f

    the

    econd-passregression atisfy:

    (21)

    Re

    =

    7

    +

    ,-

    X

    +

    h

    i

    =

    1

    This is

    equation 8)

    with

    7

    and X

    replaced

    by 7

    and X

    (the

    two-passestimates)

    nd

    replaced by

    ,

    the

    diosyncratic ricing

    rror r

    residual).

    Now

    suppose

    we con-

    sider henew set of excess

    returns,

    f

    =

    R

    -

    R',

    for

    >

    2. Given thedefinition

    f

    Rf

    and

    equation

    21),

    it follows

    that he

    sample

    mean

    of

    Rf

    is

    given by:

    (22) f

    =

    ((

    -

    O'

    X

    +

    -

    ,

    i

    =

    2,...,.

    PI is not

    ust

    any

    asset.

    It

    happens

    to be an asset

    forwhich the SDF

    beta is

    roughly

    zero

    (

    X

    =

    0)

    and the

    idiosyncratic ricing

    rror s

    very

    small

    (

    =

    0).

    Using

    these facts n

    equation

    22),

    we have

    (23)

    Rf

    =

    (3,-

    +

    h

    i

    =

    2

    ...,n.

    Now,

    since

    =

    0 it

    also follows thatthe

    average

    value

    of

    across

    i

    =

    2,

    ..,n,

    is

    roughly

    ero.

    Thus,

    the same X

    obtained

    by running

    he cross-sectional

    egres-

    sion

    for the

    P-portfolios

    an

    approximately

    it he cross-sectional

    distribution

    f

    Rfwithout constant.

    In their

    eply,

    t seems that

    V concede that

    heirmodel

    does not

    price

    the

    original

    portfolios.

    ut this

    means

    they

    have

    not dentified

    hetrue DF.

    All of the

    portfolios

    (the

    P and

    D-portfolios)

    hould

    be

    priced

    by

    the same SDF.

    Effectively

    his

    means

    there

    mustbe a

    missing

    factor hat

    prices

    PI. Since this

    factor s

    responsible

    for he

    fit f the

    original

    portfolios

    we are

    back to

    square

    one.

    Are the

    D-portfolios

    riced by

    consumption

    rowth?

    On

    thebasis

    of factor

    etas

    theanswer

    s

    clearly

    no."

    The beta matrix

    or he

    D-portfolios

    s the ame

    transfor-

    mation f

    the

    -portfolios

    sed to

    testwhether

    hey

    re

    equal

    to

    a common

    constant:

    i.e.,

    D

    =

  • 8/11/2019 Burnside 2011

    18/22

    3472

    THE MERICANCONOMIC

    EVIEW

    DECEMBER011

    consumption

    rowth,

    .47 1 fordurables

    growth

    nd 0. 186 for hemarket

    eturn. he

    SDF

    betas associated with he

    D-portfolios

    re also

    jointly tatisticallynsignificant

    (Table 10).

    Workingwith heD-portfolios lso does not alleviatethe dentificationroblem.

    The identification

    roblem

    arises

    because thereexists at least one

    nonzero

    x

    1

    vector

    x

    such

    that

    px

    =

    0,

    statistically.

    iven

    the definition f

    D

    it follows that

    Dx

    =

    0,

    statistically.

    n

    fact,

    the identification

    roblem gets

    worse,

    because the

    transformation

    is not nvertible.

    ny

    x such that

    px

    = 0

    implies

    that

    Dx

    = 0.

    But there

    may

    be additional

    x

    such

    that

    Dx

    =

    0 forwhich

    px

    0.

    This is

    hardly

    surprising,

    ince

    throwing way

    information

    s never

    ikely

    to

    improve

    dentifica-

    tion.

    Formal test tatistics

    erifying

    his re

    provided

    n the

    Appendix.

    C. The Post-BrettonWoodsEra

    My

    comment

    mainly

    iscusses

    the onclusionswe shoulddrawfrom V's evidence.

    Additional vidence from he

    post-Bretton

    Woods

    era,

    similar o the evidence ntro-

    duced

    by

    LV in

    their

    eply,

    asts furtheroubt

    n

    a

    consumption-basedxplanation

    f

    carry

    radereturns. ere I discuss evidence

    gleaned

    from

    sample

    of

    21

    developed-

    country

    urrencies verthe

    period

    1976-2010.

    The same

    sample

    of currenciess used

    by

    Burnside t al.

    (201 1)

    for he

    period

    1976-2009,

    and similar

    amples

    are used

    by

    Lustig,

    Roussanov,

    and Verdelhan

    forthcoming)

    nd Lukas Menkhoff t al.

    (forth-

    coming),

    for he

    period

    fter

    983,

    to

    study arry

    rade

    portfolios.

    My

    dataset

    consists of

    spot

    and forward

    xchange

    ratesfrom

    Reuters/WMR

    nd

    Barclays,

    available on Datastream.The raw data are

    daily

    observations f

    spot

    and

    one-month orward

    xchange

    rates. use end of month alues

    of thesedata to create

    monthly

    bservations.The data

    span

    the

    period January

    976

    to December

    2010,

    with he

    ample

    varying

    y currency.

    s in

    Lustig,

    Roussanov,

    and

    Verdelhan

    forth-

    coming),

    n each

    period,

    the

    available currencies n

    my

    sample

    are sorted nto six

    bins

    according

    o their orward iscount

    gainst

    he

    US dollar.The first in ncludes

    those currencieswith he

    smallestforward iscounts

    the

    owest nterest

    ates),

    the

    second bin the next

    smallest, tc.,

    with

    he sixthbin

    consisting

    f those

    currencies

    with

    he

    argest

    forward iscounts

    and,

    therefore,

    he

    highest

    nterest

    ates).

    then

    compute

    the

    payoff

    ssociated with

    borrowing

    ne dollar n

    order o invest

    qually

    in the

    riskless securities f the

    constituent urrencies f each

    bin. This

    procedure

    producessix currency ortfolios hat refer o as Q 1 Q2, . Q6. Then, followthe

    procedure

    dvocated

    by

    LV in

    their

    eply.

    construct he fivedifferenced

    ortfolios,

    DQ2,DQ3, ...,DQ6,

    which nvolve

    holding

    Qi

    and

    shorting

    l,

    for

    =

    2, ..,6.

    1

    measure

    payoffs

    o

    these

    portfolios

    n

    a

    monthly

    asis and

    then

    omputequarterly

    excess

    returns. o assess the

    model I use data

    for

    consumptiongrowth,

    urables

    growth,

    nd

    the market eturn

    hat re available

    from1976:11 o

    2010:1

    (sources

    for

    these data

    are described n

    Burnsideet al.

    201

    1,

    and in the online

    data

    archive).

    Using

    these

    portfolios

    find ven

    stronger

    vidence

    against

    consumption-based

    explanation

    of

    currency

    eturns. s the

    detailed results n

    the

    Appendix

    indicate,

    none of the

    betas of the

    differenced

    ortfolios

    with

    respect

    o

    consumption rowth

    and durablesgrowth s individually ignificant. he pointestimates also do not

    display

    any pattern

    f

    increasing

    with

    the nterest

    ifferential.

    ot

    surprisingly,

    s

    Table 1 1

    indicates,

    we

    cannot

    reject

    thenull

    hypotheses

    hat

    tj

    =

    0

    for

    ll

    i,

    for he

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    19/22

  • 8/11/2019 Burnside 2011

    20/22

    3474

    THE

    MERICANCONOMIC

    EVIEW DECEMBER

    011

    Table

    Factor

    etas nd

    Covariances

    or

    uarterly

    ifferenced

    ortfolios,

    Tests or pread

    nd gainst

    ero

    (A)

    Betas

    (B)

    Covariances

    Ac Ad w Ac Ad fw

    Standardrror

    ype

    Testsor o

    pread

    p- alues)

    System-OLS

    0629 0499 85

    - -

    ~

    GMM-VARHAC 0.542 0.615 0.044 0.494 0.402 0.072

    Jointestsersusero

    /?-alues)

    System-OLS

    0.763 0.618 0.016

    - -

    GMM-VARHAC 0.685 0.681 0.049 0.640 0.470

    0.133

    Notes:

    uarterly

    ata,

    976:11-2010:1.

    n

    part

    A)

    the

    egression

    quation

    s

    Reit

    a

    +

    f

    t

    -

    eit,

    here

    eit

    s he

    xcess

    eturnf

    ortfolio

    at ime

    ,

    nd

    ,

    s he

    ector

    f

    iskactorsescribed

    inTable .The

    ortfolios

    re

    Q2,DQ3,..,DQ5,

    he eturnso

    holding

    ong ositions

    n

    he

    five

    uarterlyortfolios

    2,Q3,..,Q6

    while

    olding

    short

    osition

    n he

    uarterlyortfo-

    lioQl.The ableeports- aluesorestsf he ypotheseshattj .V and tj 0 Vfor

    each actor

    .

    In

    part

    B)

    the ovariancesetweenhe xcess

    eturns

    nd

    he

    actors,

    ov

    Rehfj),

    are

    stimated

    y

    GMM. he able

    eports7-

    aluesor

    estsf he

    ypotheses

    hato

    (/f,JJ)

    =

    Cj

    V and ov

    Rehfj)

    0 Vforach actor

    .

    centers n thefact hat

    urrency

    eturns nd

    consumption rowth

    re

    approximately

    uncorrelatedwith ne another. he

    global

    financial risis s an observation

    hat uits

    LV's

    hypothesis,

    ecause it

    s also an

    episode

    in which

    US

    consumption rowth

    was

    low. The other rises

    Mexican,

    Asian, Russian,

    and

    Argentinian)

    hat

    hey

    mention

    in their

    eply

    are

    not,

    because

    they

    did

    not coincide with

    periods

    of low

    US con-

    sumption rowth.According

    to LV's

    data,

    durables

    growth

    was well above

    average

    through

    ll of the atter

    pisodes,

    while nondurable

    onsumption rowth

    was well

    above

    average during

    wo and

    slightly

    elow

    average

    n two.

    Overall, however,

    there s little evidence

    that

    average

    carry

    trade

    returns an

    be

    explained by

    increased

    exposure

    to stock market isk

    during

    periods

    of

    reces-

    sion,

    market

    ownturns,

    r

    market

    olatility.

    o

    demonstrate

    his,

    first

    egress

    he

    monthly

    eturns o the

    DQ6

    portfolio

    equivalent

    o

    LV's HML

    carry

    rade

    portfolio)

    on the

    monthly

    xcess returns f

    the

    value-weighted

    US stock market

    the

    factor).

    The

    market eta of

    DQ6

    in the

    period

    1976:2-2010:12 is 0.18 and

    s statisti-

    cally significant.

    s in

    the

    quarterly ample,

    thisbeta is

    much too small

    (by

    a

    factor

    ofmore thanfive)toexplaintheaveragereturn f thecarry rade.

    I

    then

    divide

    my

    data into

    recessions

    and

    expansions

    as

    defined

    by

    the

    NBER,

    periods

    of

    high

    and low stock

    market

    olatility,

    nd

    periods

    of

    high

    and low

    stock

    market eturns.

    o measure

    volatility,

    use

    the

    daily

    standard eviation

    of the mar-

    ketexcess returnn

    each

    month.

    Months

    n

    which this

    measureof

    volatility

    s

    more

    than one

    standard

    deviation above

    its mean are

    denoted as

    "high

    volatility."

    o

    divide

    thedata into

    periods

    of

    high

    and low

    stock

    returns,

    denote

    months

    n

    which

    the

    market xcess

    returns

    more than ne

    standard

    eviationbelow its

    mean as "low

    return"

    months. he

    market eta

    of

    DQ6

    is

    0.14

    during xpansions

    and 0.26

    during

    recessions. t is

    0.25 in

    periods

    of

    highvolatility

    nd

    0.

    1 1

    in other

    eriods.

    t is 0.33

    in periodsof low stockreturns nd 0.15 in otherperiods.In each case, however,

    the

    difference

    etween the

    two betas is

    not

    statistically

    ignificant,

    nd

    the betas

    certainly

    emain

    nsufficientlyarge

    to

    explain

    the

    average

    returns f the

    carry

    rade.

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    21/22

    VOL.

    01

    NO.

    BURNSIDE:ISK

    REMIA

    ND

    ONSUMPTION:OMMENT

    3475

    Table 2 SDFBetas or

    uarterly

    ifferenced

    ortfolios,

    Tests or preadnd gainst ero

    Model

    i)

    Model

    ii)

    Model

    iii)

    Standardrrorype Testsor o preadp- alues)

    System-OLS

    0.333 0.185 0.273

    GMM-VARHAC

    0.280 0.103

    0.301

    Joint

    estsersusero

    p- alues)

    System-OLS

    0409

    068 0.399

    GMM-VARHAC

    0.332 0.139 0.431

    Notes:

    uarterly

    ata,

    76:11-20

    0Q:I.

    he

    egressionquation

    s

    Reit

    ai

    -

    mtimeit,

    where

    t

    1

    -

    (f,

    f)'b,

    s

    he

    ample

    ean

    f

    ,,

    nd he ectortakesn ne f he ol-

    lowing

    hreealues:

    i)

    b

    =

    (-21.0

    129.9

    .46)'

    corresponds

    o V's

    wo-pass

    stimateith

    a

    constant),ii)

    =

    (37.0

    4.7

    .65)' corresponds

    o V'sGMM stimateith o

    onstant),

    and

    iii)b (6.74

    3.3

    .31)' the

    alibrated

    odel).

    ere

    eit

    nd

    ,

    re he eturnsnd ac-

    tors,

    escribedn he oteo able The able

    eports-values

    orestsf he

    ypotheses

    hat

    m ^ia aAm=^i-

    V. Conclusion

    To

    explain

    cross-sectional variation

    in

    expected

    returns,

    risk-based

    story

    requires

    that ssets withnonzero

    returns e correlatedwith the

    proposed

    SDF. As

    Section

    II

    demonstrates, owever,

    LV's

    consumption-based

    DF is

    approximately

    uncorrectedwith ll of thereturns

    hey tudy.

    Given this

    fact,

    ne cannot

    reject

    hat

    LV's estimates re consistent

    with

    consumption

    isk

    explaining

    none of the

    cross-

    sectionalvariation

    n

    the

    expected

    returns

    hey

    tudied.

    I

    have

    argued

    hat he tatistical

    nsignificance

    f the SDF

    betas eads to two addi-

    tional

    problems

    with

    LV's conclusions.

    First,

    t

    mplies

    that ne

    cannot

    gnore

    am-

    pling

    uncertainty

    n

    the

    betas when

    conducting

    nference

    bout factor isk

    premia.

    The statistical

    ignificance

    V

    point

    o

    argely

    vanisheswhen standard rrors

    ppro-

    priately

    eflect his

    uncertainty.

    econd,

    the

    degree

    of

    uncertainty

    bout the betas

    implies

    that

    hefactor isk

    premia

    re

    veryweakly

    dentified.

    his makes

    asymptotic

    inference ess reliable.

    Confidence ets

    that re robust o weak identification

    uggest

    thatLV's data

    are

    approximately

    ninformativebout

    consumption

    isk.

    Finally,

    have

    argued

    that

    LV are able to

    report

    trikingly

    igh

    R2 measures of fit

    because theirmodel

    includes a constant

    ricing

    rror,

    which

    s treated s

    part

    f the

    model's predicted xpectedreturns.When this onstants excludedfrom hemodel,

    the

    R2 statistics re much

    smaller

    nd,

    n

    many

    cases,

    negative.

    A central

    oint

    of

    my

    discussion

    s that he betas

    of

    consumption

    actors re

    very

    poorly

    stimated.

    his is

    why

    a

    consumption-based

    model is difficult

    o

    rejectusing

    a

    formal est f the

    pricing

    rrors.

    f

    a

    great

    deal moredata

    were

    collected,

    one

    might

    obtain

    ufficientlyrecise

    estimates

    f thebetas

    to enable

    sharper

    onclusions

    about

    the

    model. But with

    he data we

    have,

    both LV's data

    and

    highfrequency

    ata

    from

    the

    post-Bretton

    Woods

    period,

    there

    s no

    justification

    or

    drawing strong

    on-

    clusion

    in

    favor

    f a

    consumption-based

    model. Burnside

    et al.

    (2011)

    have

    argued

    that

    out-of-sample

    isk s a

    potential xplanation

    for the

    returns o

    the

    carry

    rade.

    If thesecurrentlyut-of-sampleventsoccur nthefuture,tmaywell turn ut that

    a

    consumption-based

    model works.

    At the

    moment,

    however,

    model

    based

    on in-

    sample

    variation

    n

    consumption

    oes

    not work.

    This content downloaded from 143.107.200.198 on Thu, 14 Aug 2014 21:01:37 UTCAll use subject to JSTOR Terms and Conditions

    http://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsphttp://www.jstor.org/page/info/about/policies/terms.jsp
  • 8/11/2019 Burnside 2011

    22/22

    3476

    THE MERICANCONOMIC

    EVIEW DECEMBER

    011

    REFERENCES

    Burnside,

    raig.

    010. Identification

    nd nferencen Linear tochasticiscount actor odels."

    Nationalureau fEconomicesearch orkingaper 6634.Burnside,

    raig.

    011."The Cross ection f

    Foreign

    urrency

    iskPremiand

    Consumption

    Growth isk:Comment:ataset." mericanconomic eview,

    ttp://www.aeaweb.org/articles.

    php?doi=10.

    257/aer.017.3456.

    Burnside,

    raig,

    Martn

    ichenbaum,

    saac

    Kleshchelski,

    nd

    Sergio

    ebelo. 011. Do PesoProb-

    lems

    xplain

    he eturnso he

    arry

    rade?" eview

    f

    inancialtudies

    24(3):

    853-91.

    Center

    or esearchn

    Security

    rices. 010. CRSP

    Monthly

    reasury

    S Database uide."

    ttp://

    www.crsp.com/documentation/pdfs/monthlyreasury.pdf.

    Cochrane,

    ohn

    . 2005.

    Asset

    ricing.

    nd .

    Princeton,

    J: rinceton

    niversity

    ress.

    den

    Haan,

    Wouter

    .,

    ndAndrew. Levin. 000. Robust ovariance atrixstimationith ata-

    Dependent

    AR

    rewhitening

    rder." ational ureau fEconomic esearchechnical

    orking

    Paper

    55.

    Fama,

    ugene

    .,

    ndKenneth.

    French.993. Common

    isk

    actors

    n he

    ReturnsnStock nd

    Bonds." ournal

    f

    inancial conomics

    33(1):

    3-56.

    Fama, ugene ., ndJames . MacBeth.973. Risk, eturn,ndEquilibrium:mpiricalests."

    Journal

    f

    olitical

    conomy

    81(3):

    607-36.

    Hansen,

    arsP. 1982.

    Large ample roperties

    f

    Generalized ethod fMoments

    stimators."

    Econometrica

    50(4):

    1029-54.

    Jagannathan,

    avi,

    nd

    ZhenyuWang.

    998.

    An

    Asymptoticheory

    or

    stimating

    eta-Pricing

    Models

    sing

    ross-Sectional

    egression."

    ournal

    f

    inance,

    3(4):

    1285-1309.

    Kan,

    Raymond,

    ndChu

    Zhang.

    999a. GMM ests f tochasticiscount

    actor odels ith

    se-

    lessFactors."

    ournal

    f

    inancial

    conomics

    54(1):

    103-27.

    Kan,

    Raymond,

    nd

    Chu

    Zhang.

    999b.

    Two-Pass

    ests fAsset

    ricing

    odelswith

    seless ac-

    tors." ournal

    f

    inance,

    4(1):

    203-35.

    Lustig,

    anno,

    ikolai

    oussanov,

    nd

    Adrien

    erdelhan.

    orthcoming.

    Common

    isk actorsn

    Currency

    arkets."eview

    f

    inancialtudies.

    Lustig,

    anno,

    ndAdrien

    erdelhan.007. The

    ross ectionf

    oreign

    urrency

    isk remia

    nd

    Consumptionrowthisk." mericanconomiceview97(1):89-1 7.

    Menkhoff,

    ukas,

    ucio

    arno,

    Maik

    chmeling,

    ndAndreas

    chrimpf.

    orthcoming.

    Carry

    rades

    and

    Global

    oreignxchange

    olatility."

    ournal

    f

    inance.

    Shanken,

    ay.

    992. On

    he stimationf

    Beta-Pricing

    odels."

    eview

    f

    Financialtudies

    5(1):

    1-33.

    Wright,

    onathan

    . 2003.

    Detecting

    ack f

    dentification

    nGMM."

    conometric

    heory

    19(2):

    322-30.

    Yogo,

    Motohiro.

    006.

    A

    Consumption-Based

    xplanation

    f

    Expected

    tock

    eturns."ournal

    f

    Finance

    61(2):

    539-80.