The Value of a Good Repution Online (Application to Art Auctions)

8/2/2019 The Value of a Good Repution Online (Application to Art Auctions)

1/34

SELLING ONLINE VERSUS LIVE

EIICHIRO KAZUMORI AND JOHN MCMILLAN

Abstract. A seller choosing between auctioning online and livefaces a tradeo: lower transaction costs online against more rentsleft with the bidders. We model this tradeo, and apply the theoryto auctions of art. The crucial parameter for whether the seller doesbetter online than live is not the expected price but the valuationuncertainty.

1. Introduction

Online auction sales reached $13 billion in the United States in 2002,still growing at 33 percent per year. Cut owers, seafood, classic cars,

jewelry, coins, antiques, and art are now auctioned online as well as intraditional live sales. In what follows we model the s ellers choice ofsale venue, online versus live, and apply the theory to auctions of art,where prices and valuation uncertainties are extreme.

The internet has made markets more ecient by making it easier forbuyers to learn what is available where for how much. The lowering

of these search frictions has made pricing more competitive, as vari-ous empirical studies have found (Baye, Morgan, and Scholten, 2001,Brown and Goolsbee, 2002, Brynjolfsson and Smith, 2000, Clay, Kr-ishnan, and Wol, 2001, Ellison and Ellison, 2001). There is one set offrictions that the internet has not reduced, however: those of verifyingquality. Such frictions mean, we will argue, that online prices tend tobe less competitive.

Are there limits to what can be sold online? Industry wisdom saysthe internet is not suitable for selling high-end items, because potentialbuyers will resist bidding large sums for goods they have not seen.Counterexamples exist, however. The town of Bridgeville, California

Date: January 29, 2003.Key words and phrases. auctions, electronic commerce, art market.California Institute of Technology; [email protected]. Stanford

University; [email protected]. We acknowledge support from theCenter for Electronic Business and Commerce at Stanford Universitys GraduateSchool of Business and from a Stanford dissertation fellowship, and thank Yee ManChan, Brian Poi, and Ha Quach for research assistance.

1


2/34

2 EIICH IRO KAZUMORI AN D JOHN M CMILLAN

was auctioned on eBay for $1.8 million. Most of the 249 bidders madetheir oers based solely on digital photographs and descriptions of the

town on a Web site, according to a news report.1

A Ferrari sold oneBay Motors for $330,000. A 1909 baseball card sold on eBay for $1.3million. A rst printing of the Declaration of Independence, from July4, 1776, sold on Sothebys.com for $8.1 million.

There are two main dierences between online and live auctions.First, less information is available for online bidders. In a live auction,bidders inspect the item at the preview and (in the case of art auctions)can get advice from the auction-house experts present. In an onlineauction, bidders get only what information is posted on a web page.The fuzziness of pictures on a computer screen means an online biddersinformation is limited.

Second, transaction costs are higher in live auctions. Online, thesellers costs are those of running a web site. Live, the auctioneerpays the costs of mounting the pre-sale exhibition and in running thetheatrical performance that makes up a live auction. The live sellermust wait until a suitable auction is scheduled, bringing costs of delay.Bidders have negligible participation costs online, whereas to bid livethey incur travel and other participation costs.

In the auctions of tuna at Tokyos Tsukiji market, where a singlesh goes for up to $15,000, buyers spend half an hour examining thetuna to assess their suitability for sashimi and sushi. This adds up tohundreds of person-hours of transaction costs in these live auctions.

In online seafood auctions in Europe, by contrast, the auctioneer usesa letter-grade system to stipulate the freshness of the sh, thus econ-omizing on the bidders time, though possibly at the cost of inferiorinformation. When ower auctions in the Netherlands went online,prices fell ve percentarguably because the bidders found it hard toassess the owers quality from an on-screen picture.2

Any seller choosing between auctioning online and live, then, facesa tradeo. On the one hand, the costs to the seller and the biddersof participating are lower online. On the other hand, online bidders,worrying more about the winners curse, bid lower. We will model thetradeo of lower transaction costs online against more rents left with

the bidders. The winning bidder does better in an online auction thanin a live one in our model, despite being more uncertain about the

1San Jose Mercury News, December 28 2002, p. 1A. T he $13 billion gure citedearlier comes from CNET News.com, October 18, 2002.

2On the Tsukiji sh auctions, see McMillan (2002, p. 65); on t he online Euro-pean sh auctions, www.pefa.com/pefaportal/en/index.htm; on the Dutch owerauctions, Koppius, Van Heck, and Wolters (2002).


3/34

SELLING ONLINE VS. LIVE 3

assets value (because the competitors also bid lower). The gains canbe mutual: in certain cases the seller, too, does better online (because

of the saving in transaction costs).What are the limits of online auctions? The common belief among

online sellers seems to be that it is the items expected price that de-termines whether selling online works better for the seller; high-valueitems should be sold live. Our theory suggests, to the contrary, thatthe crucial variable is not the expected price but the extent of valua-tion uncertainty. Valuable items can be successfully auctioned onlineif their variance of value is low.

2. Art Auctions

Internet auctions of art are nearly as old as electronic commerce.

In 1995, Donna Rose, a Beverly Hills art dealer, ran some on her sitewww.artbokerage.com (Turban, 1997). Now, many sites oer originalartworks at xed prices or by auction. The biggest player is Sothebys,which in 2000 revamped its time-honored sale mechanism and startedauctioning online via Sothebys.com.

An online seller of heterogeneous goods of any kind faces questionsof quality. The success of eBay, for example, is in part due to itsseller-reputation mechanism, reassuring bidders by tracking previousbuyers experiences (Baron, 2001, Houser and Wooders, 2000, Lucking-Reiley, Bryan, Prasad, and Reeves, 2000, McDonald and Slawson, 2000,Resnick and Zeckhauser, 2001). More than anything else that is sold

online, art is prone to vast uncertainties about authenticity and merit.Art thus tests the scope of online sales. Nevertheless, art collectorsreadily adapted to bidding online. Shortly after online art sales began,Business Week noted, "what is astonishing many experts is how quicklycollectors have accepted buying expensive works over the Net withoutseeing them," and quoted Craig Moett, president of Sothebys.com,saying, "That we would be selling works in the $20,000, $30,000, and$40,000 range is a surprise."3

Sothebys online auctions have a checkered history. Sothebys.combegan operations in January 2000 as a joint venture between Sothebysand Amazon.com. This alliance was dissolved after nine months; theend came ve days after Sothebys pleaded guilty to colluding withChristies to x fees in their live auctions. Sothebys operated theonline auctions alone from October 2000 to June 2002. During thistime it sold goods worth over $100 million in a period when the artmarket in general was depressed, but lost money because of the web

3Business Week, July 3, 2000, p. 152.


4/34


sites high set-up costs.4 Then it formed a joint venture with eBay.The joint venture replaced eBay Premiere, eBays own site for high-

end auctions, while Sothebys online business moved to eBays web site,allowing Sothebys to cut sta and expenditures in its online division.

Live auctions at Sothebys have four steps: consignment, catalogu-ing, exhibition, and sale. After a prospective seller contacts Sothebys,its experts check the items authenticity and appraise its value. Theowner then consigns the item to Sothebys, which puts it in an upcom-ing auction. Sothebys auctions have a theme, such as 20th CenturyWorks of Art or Egyptian, Classical, and Western Asiatic Antiques,so the sale waits until there is an auction that it ts. The auctioncatalogue, available about one month ahead of the auction, containsa description of the item, its history, reference notes, and an upper

and lower estimated s elling price. The estimate is "an approximatevaluation, based whenever possible on comparable auction values," ac-cording to Sothebys (Hildesley, 1984, p. 57). A few days before theauction, Sothebys mounts an exhibition of the items to be sold. Theexhibition is open the public and Sothebys specialists are on hand toanswer questions.5 Finally there is the auction itself, run with openascending bidding. Any major sale is a social event for the glitterati.Space in the auction room is scarce so it is rationed (Watson, 1992, p.11), desirable seats being assigned to clients who have spent lavishlyin the past.

Sothebys online auctions dier from its live auctions in several re-

spects. First, Sothebys is the consignor in only a fraction of Sothe-bys.com auctions. It has formed partnerships with about 5,000 inter-net associates, independent dealers who oer goods for sale at Sothe-bys.com. The associates authenticate their own goods and make theirown appraisals. Second, rather than large auction events with specicthemes at discrete times, miscellaneous items are continuously on oervia the web site. (At any time the site lists around 13,000 items forsale.) Third, there is no printed catalogue and no exhibition; the de-scriptions are solely online. Fourth, the auction is run by eBays rules;in particular, each auction has a xed end-time (and so there is the

4Sothebys incurred costs of $60 million in the rst two years of its online oper-ation, mostly in set-up costs (Wall Street Journal, January 31, 2002, p. B4).

5It is buyer beware. A catalogue states, Neither we nor the Consignor make anywarranties or representations of the correctness of the catalogue or other descrip-tion of physical condition size, quantity, rarity, importance, medium, provenance,exhibitions, literature or historical relevance of t he property . . . Prospective bid-ders should inspect the property before bidding to determine its condition, size orwhether or not it has been repaired or restored (Sothebys, 1988).


5/34


possibility of last-minute bidding, analyzed in eBay auctions by Bajariand Hortacsu, 2002a, and Ockenfels and Roth, 2002). Unlike eBays

regular auctions but like Sothebys live auctions, however, the onlineart auctions come with expert appraisals.

Sothebys main cost of running online auctions is in producing theweb-site descriptions. This costs no more, presumably, than print-ing the catalogue for a live auction. The other live-auction costs areavoided. For bidders, participation is easier online. The live bidderstransaction costs go beyond travel and other pecuniary costs. SothebysLondon and New York salerooms have an elitist ambiance, said to bedaunting to newcomers. This world has been too intimidating and rar-ied, said Sothebys president Diana D. Brooks when Sothebys.comwas starting up. We want our site to be a friendly place to the poten-

tially millions of people whove never bought art before. In the rstfew months of their online auctions, according to Sothebys, 85 percentof the bidders had never bought anything at Sothebys before.6

An online bidder sees a photograph of the item for sale, sometimesa paragraph of biography of the artist, a terse statement of the workscondition, and a lower and upper estimate. Figures 1 and 2 exem-plify how muchor how littleinformation the bidders are given tohelp them assess the artworks value. (To see more such examples, goto http://sothebys.ebay.com and click on a few of the listings.) Aphotograph does not usually do justice to a piece of art.

Is the internet right for selling art? Sothebys move to online auc-

tions was greeted skeptically. Christies, Sothebys rival in the globalart duopoly, has deliberately stayed away from online auctions. An un-convinced art dealer remarked, What we sell is something that needsto be looked at and discussed. The chief executive of the traditional artauctioneer Butterelds, Geo Iddison, said, Buyers expect to touchthe works of art in person.7 Next we develop a model that both sharp-ens and amends the observation that potential buyers prefer to see theart before bidding.

3. The Model and Equilibrium

The decision makers: We use the common-value model: that is, there

is a single item for sale with an objective value to the potential buyers,dened perhaps by its future resale value, which at the time of bidding

6Quote from Economist, January 27, 2000. New-bidders number from Worth,Winter 2000/2001, www.worth.com.

7The rst quote is from a L ondon oriental art dealer, Giuseppe Eskenazi, inEconomist, January 27, 2000. Iddison quote: San Jose Mercury News, Aug. 2,2002.


6/34


no one knows. The winning bidder gets an ex post payo of s p,the dierence between the common value s and the price p. The seller

has zero value for the asset if unsold. A bidder i receives a scalarsignal xi about the assets value s. The expected value E[sjx], wherex = (x1;:::;xn), is nondecreasing in each argument. The potentialbuyers and the seller are risk neutral and share a common prior on sand x:

The fact that art auctioneers announce a presale estimate indicatesthe existence of some underlying common value and implies the alter-native model, independent private values, does not t (though valuescould, more generally, be aliated in the sense of Milgrom and Weber(1982a).) Renoir apparently believed the common-value model appliesto art. "Get this into your head, no one really knows anything about

it," he said. "Theres only one indicator for telling the value of paintingsand thats the sale-room" (quoted by Duret-Robert, 1978). Renoirswords "no one really knows" and "the value" invoke the common-valuemodel, as does his assertion, la Wilson (1977), that price revealsvalue.

An online auction in our model is a standardthat is, frictionlessascending auction. A live auction has endogeneous entry of bidders; theseller and the bidders incur transaction costs. There are N potentialbidders. In an online auction, with zero participation costs, all N areactual bidders. In a live auction, with positive participation costs, asubset of n of them (with n N) are actual bidders.

The decision process: Nature sets the common value s. Neither theseller nor any of the bidders learns the realization ofs. Then the sellerdecides whether to sell the asset in a live auction or online.

In an online sale:

The seller reveals information about the asset on its webpagetruthfully, because of its guarantee8 and the eBay feedback rat-ing system.

Bidder i obtains the signal xoni from the distribution Fon.

Bidders compete in an ascending auction, with a reserve priceron set by the seller.

In a live sale, on the other hand: The seller pays a cost cs of running the live auction.

8"Each seller guarantees that the authorship, period, culture or originof the lot is as set out in the Guaranteed sections of the View Itempage in the description of the lot;" this guarantee is valid for three years.http://pages.sothebys.ebay.com/help/rulesandsafety/guarantee.html


7/34


The seller publishes a description of the asset in an auctioncataloguetruthfully, for reputational and legal reasons.9

Each potential bidder decides sequentially whether to enter theauction. At the time of this choice, each learns the number ofbidders who have already entered.10

Each bidder who enters pays a transaction cost (to attend thepreview and inspect the asset, for example) of cb:

After deciding to enter, each bidder receives a signal xoffi fromthe distribution Foff.

The bidders compete in an ascending auction with a reserveprice roff.

The bidders information: Each of the signals Xoffi and Xoni belongs

to a mean-dispersion family with a mean-zero base random variable Z;

with a distribution function FZ and continuously dierentiable densityfZ and with mean and dispersion parameters (; off) and (; on).This implies

Fon(x) = FZ(x

on); Foff(x) = FZ(

x

off)

A member of a mean-dispersion family is obtained by a shift in themean and the dispersion of the base distribution. (For example, ifX is normal, the base random variable is N(0, 1).) Many distributionsbelong to the mean-dispersion family in addition to normal, such as thelognormal, Poisson, uniform, and extremum-value distributions. The

signals (Xj1;:::;X

j

N ), j = on, o are independent.We assume what we will call the signal-sensitivity condition: thesensitivity of expected value to signal does not vary excessively withthe level of the signal: that is, S and Xj; for j = on, o, are such thatthere exists m > 0; M < 1 such that for all x;

(3.1) m @

@xiE[sjX = x] M

(Recall that E[sjX = x] is nondecreasing in each argument.) Forexample, if m = M, a $1 increase in the signal increases the biddersexpected value by the same amount regardless of the level of the signal.

9A live auction catalogue states, "We guarantee the authenticity of Authorshipof each lot contained in this catalogue . . . Authorship, locations, the identity ofthe creator, the period, culture, source of origin of property, as the case may be, asset forth in the Bold Type Heading of such catalogue entry" (Sothebys, 1988).

10This is as in McAfee and McMillan ( 1987). Levin and Smith (1994) modeledsimultaneous entry. The two formulations are s imilar, since both are driven by thebidders zero-prot conditions. Other models of entry include McAfee and Vincent(1992), Hausch and Li (1993), McAfee, Quan, and Vincent (2002), and Ye (2001).


8/34


(m = M holds, for example, in the mean-common-value model, inwhich the value s equals the average signal, P xi=n.)

We also assume that the hazard rate (that is, f(x)=[1 F(x)]) isincreasing for both of Xoffi and X

oni . Note that, if Yi;n is the ith

highest realization out of n iid samples, then (from David, 1981, p.129):

E(Yoffi;n ) = off + offE(Yzi;n); E(Y

oni;n ) =

on + onE(Yzi;n):

The crucial parameter in our model, the inspection parameter k,represents the extent to which bidders are better informed about theassets value in a live auction than an online one. The live auction,we assume, provides a more precise signal: off = kon for some given

0 6 k 6 1. The parameter k measures the lower dispersion in liveauctions than online; k = 0 means a complete elimination of dispersionfrom seeing the item live, and k = 1 means no reduction of dispersion.

The rst result identies the equilibrium.Lemma 1: In an online auction without a reserve price, each bid-

ders expected prot is,

onB =

ZHon(xi; N)(1 F

on(xi))dxi

where

Hon(xi) ZXi

@vi(x; x

i)@x jx=xif(xijxj xi; j 6= i)dxi,

vi(x) = E[sjX = x]:

The sellers expected prot is

onS = Es N

ZHon(xi; N)(1 F

on(xi))dxi:

In a live auction, the seller sets the reserve price equal to zero, andeach bidders expected prot is zero. The sellers expected prot is

off

S = Es noff

cb cs

= Es noffZ

Hoff(xi; noff)(1 Foff(xi))dxi cs:

The number of bidders noff is determined endogenously byZ

Hoff(xi; noff)(1 Foff(xi))dx = cb:


9/34


The proof is in the appendix. Recall the concept of marginal revenueintroduced by Bulow and Roberts (1989). In our setting, this is

MRi(xi) =1

f(xi)

d

dxi[vi(xi;xi)(1 F(xi))]:

(We suppress dependence on n unless needed.) By our assumption, abidders marginal revenue is monotone increasing in own signal. Thusthe seller sells the asset to the bidder with the highest signal. Sincewe assume independent signals, the revenue equivalence theorem holds(Bulow and Klemperer (1996)), so our results for online auctions applynot only to ascending auctions but also to other standard formats likethe rst-price and second-price sealed-bid auctions.

In a live auction, the participation cost determines the number of

bidders. Since increasing the number of bidders decreases the expectedprot each gets, the number of bidders is established (up to integerconstraints) by a zero-prot condition (McAfee and McMillan, 1987).The n bidders who enter share a total expected prot ofEs Ep, theexpected dierence between the assets value and the winning bid. Exante, they each get (1=n)th of this, or (EsEp)=n: Bidders enter untilthis expected prot is driven down to their participation cost cb; thus(EsEp)=n = cb. Therefore, Ep = Esncb, which says that, ex ante,the price is the assets value minus the sum of the bidders participationcosts. The cost of participating limits the number of bidders who enterand weakens the bidding competition. The seller indirectly bears, via

a lower price, the total bidder participation costs.The tradeo between online and live auctions that, in a live auction,

the number of bidders, n; can be less than the number of bidders in anonline auction, N, implying less competition live than online. In thelive auction, however, the information available to the bidders, Xof f;is more precise than in the online auction, Xon, tending to increase thecompetitiveness of the bidding.

The dispersion of value estimates is a crucial determinant of thechoice of online versus live sales. Next we examine the eects on pay-os of varying the dispersion parameter. These results serve as a pre-liminary for the analysis in the next section and are of some theoretical

interest by themselves. We rst examine online auctionsthat is, stan-dard auctionsin terms of dispersion parameters, then live auctionsthat is, auctions with endogenous entry. Previous results, like Milgromand Weber (1982b) and Athey and Levin (2001), are order-based. Bycontrast, this approach, by assuming the signal-sensitivity condition(3.1), provides a quantitative prediction of expected prots as a func-tion of dispersion.


10/34

10 EI ICH IRO K AZ UMORI AN D JO HN M CMILLAN

Proposition 1: In an online auction, for any 1 and 2 such that1 (m=M)2, each bidders expected prot is lower and the sellers

expected prot is higher with 1 than with 2.An intuition for this is as follows: as falls, the bidders have a

more precise estimate of the asset, so they bid more aggressively. As aresult, the sellers expected prot rises and the bidders expected protdecreases. A more quantitative intuition is that, the bidders expectedprot is a function of the dierence between the expected value of thehighest signal and the second highest signal. This dierence is, by theformula (EY1;n EY2;n) = (EZ1;n EZ2;n), monotone decreasing in.

Proposition 2: In a live auction, for any 1 and 2 such that1 (m=M)2, the number of bidders is smaller and the sellers prot

is higher with 1 than with 2.The argument is straightforward: when the signal is less accurate,

bidders bid more aggressively, so a bidders expected prot increases.Anticipating this, more bidders enter the auction. The added biddersdo not, however, increase the sellers revenue: the zero-prot conditionmeans the sellers payo is the value of the asset minus the biddersentry costs.

Payos are roughly monotonic in variance, according to these twopropositions. In the special case of m = M (which means, recall,that the sensitivity of expected value to signal does not vary with thelevel of the signal), they are actually monotonic. In either a live or

an online auction, the seller prefers there to be a smaller valuationuncertainty. In an online auction, bidders prefer a wider valuationuncertainty, b ecause it means weaker bidding competition. In a liveauction, however, the eect of increasing valuation uncertainty on thebidders payos is ambiguous: the increased winners curse eect tendsto lower the bids, but the increased variance induces extra bidders toenter, creating more competition and higher bids.

The expected value of the asset aects the sellers choice betweenonline and live auction less than the variance. To illustrate this, we goto temporarily to a special case of our model.

Proposition 3: In both live and online auctions, assuming the

mean-common-value model (that is, s =P

xi=n), the number of bid-ders and the bidders expected prot are independent of .

The intuition is that bidders payo is determined by the relativecompetition with other bidders. As a result, a shift in the items ex-pected value, common to all bidders, does not aect the bidders pay-os. The seller gets all the benets of any increase in the expectedvalue. Thus, in the case of the mean-common-value model, an increase


11/34


in the expected value of the asset, with the variance and other para-meters held constant, would not induce the seller to switch between

online and live sale.

4. Live versus Online

Our main result shows that the choice between selling live and onlinedepends upon three paramaters: , the items expected value, , theestimate variance, and k, the extent to which this variance is lower livethan online.

Proposition 4: There exists k < 1 such that, for all k k; thereexist ; such that, for all and , the sellers expectedprot from selling live is higher than that from selling online.

The intuition is that, if the asset has high valuation risk, then thebenet from information revelation is higher in a live auction; in ad-dition, the dierence in the number of bidders between the live andthe online auction is smaller. Thus the seller prefers the live auction.Another way of saying this is that the transaction cost is lower relativeto the value of the asset. In a live auction, the s eller has to pay notonly her own transaction costs but also, indirectly, the bidders partic-ipation costs. If the expected price from a live auction does not coverthese transaction costs, then the seller will not sell live.

Even for a highly valuable item, an online auction can yield more

for the seller than a live one. The proof of Proposition 4 reveals threeeects that push for selling online: the transaction costs cb and cs beinghigh enough, the inspection parameter k being high enough, and theestimate variance being low enough.

For an asset with a given mean and variance, an online auction gen-erates the higher return for the seller if the inspection parameter k isclose enough to one, meaning the online bidders information is nottoo far inferior to the live bidders information. It is the nature ofthe item for sale that determines k. Thus the usable interpretation ofProposition 3 is that, for categories of assets with a given k parameter,online sales are justied for assets whose variance of value is not too

high. High-value items can be sold online if k is close enough to one.To see the role of the inspection parameter, consider how art diersfrom nancial assets. For a piece of art, the auction preview reducesthe valuation risk (that is, k < 1). By contrast, staring at a 10,000 yennote will not tell you anything about the risk of holding yen (that is,k = 1). As a result, there is no merit in trading nancial assets live,given the saving in transaction costs from online trading.


12/34


The force of the parameters k and suggests that the limits of onlineauctions dier for dierent categories of assets. A Rembrandt painting

reproduced on a computer screen loses much of its vigor: the depth andtexture of the oil paint, the gradations of light and shade, the subtletiesof the colors. For oil paintings, therefore, k is likely to be close to zero,so online bidders would bid cautiously to avoid the winners curse, andlive auctions would tend to yield better returns for the seller. Similarlywith an antique piece of clothing, such as a ninteenth century Chineserobe: it is hard to see on-screen what condition the fabric is in. Withprints, photographs, coins, and stamps, arguably, something closer tofull information can be conveyed to online bidders, so k is closer to oneand the limits of online auctions are reached at higher expected values.

The dispersion also varies across categories. A van Gogh painting

being unique, it is hard to extrapolate from past auction prices of othervan Goghs and so its is high. A Chagall print, by contrast, may existin fty or more identical copies, several of which could be auctioned ina given year (Pesando, 1993). With a nonunique item like printora photograph or a coin or a stampthe history of prices fetched byothers just like it is useful in assessing its current value, so its islower than with a unique item like a painting.

The record high online price, $8.14 million for an original of the Dec-laration of Independence at Sothebys.com in 2000, exemplies these ef-fects. Other copies exist, but they are rare (there are 25 known copies,but only four are in private hands), so there was little transaction

history. Its estimate dispersion was therefore quite high: Sothe-bys.coms low and high estimates were $4 million and $6 million.11

Being a printed document, however, its inspection parameter k wasclose to one, arguably, so selling online rather than live entailed littlereduction in bidding competition.

We now provide two numerical examples. Three bidders are inter-ested in an item on oer; its value is the mean common value,

Pxi=3.

The expected value, the inspection parameter, and the participationcost are the same in each case; the only dierence is the estimate vari-ance. Nevertheless, it turns out that in one case the seller chooses alive auction; in the other, online.

Example 1 : In an online auction, the bidders signals are distrib-uted uniformly on [2,000, 18,000], based on the information the sellerprovides via the internet. Instead, the seller could hold a live auction,in which information is 75 percent more accurate, in that the bidders

11Report dated January 11, 2000 on on www.auctionwatch.com.


13/34


signals are distributed uniformly on [8,000, 12,000]. In the live auc-tion, the seller has zero transaction costs but each bidder pays $100 to

participate.Solution: First, we compute the seller and bidder prots from the

online auction. Consider a symmetric equilibrium where a bidder dropsout when the bid price reaches the assets expected value given thesignals of the bidders who have already dropped out and assumingall remaining bidders have the same signal as that bidder. In expec-tation, bidder 1 has the signal $14,000, bidder 2 $10,000, and bid-der 3 $6,000. Bidder 3 drops out at $6000. Bidder 2 drops out at[6,000+(10,0002)]/3 = $8,667. The expected price is $8,667 and eachbidders expected prot is (10,000-8,667)/3 = $433. Second, we com-pute the prots from the live auction. In expectation, bidder 1 has

the signal $11,000, bidder 2 $10,000, and bidder 3 $9,000. Suppose allthree bidders enter. Bidder 3 drops out at $9,000. Then bidder 2 dropsout at [9,000+(10,0002)]/3 = $9,667. The sellers expected revenue is$9,667. Each bidders expected prot is 334/3 = $111. This is higherthan the entry cost of $100, so all three enter. Third, the sellers deci-sion: online, the expected return is $8,667; live, it is 9,667-(3100) =$9,357. Thus the seller holds a live auction.

Example 2: As in example 1, except for the distributions: online,the bidders signals are distributed uniformly on [9,200, 10,800]; live,the dispersion is reduced 75 percent, to [9,800, 10,200].

Solution: First, we compute the prots from the online auction. On

average, bidder 1 has the signal $10,400, bidder 2 $10,000 and bidder3 $9,600. Bidder 2 drops out at (9,600+20,000)/3 = $9,867. Thus theexpected price is $9,867 and each bidders expected prot is 133/3 =$43. Second, we compute the prots from the live auction. On average,bidder 1 has the signal $10,100, bidder 2 $10,000 and bidder 3 $9,900.Bidder 3 drops out at [9,900+(10,0002)]/3 = $9,967. The expectedprice is $9,967 and the expected prot of any bidder is 34/3 =$11.Given this prot, the third bidder does not enter. Will the other twoenter? Now the expected price is $9,933, and each bidders expectedprot is 67/2 = $34. Thus the second bidder does not enter. Only oneenters, and bids some small amount " (depending on the formulation

of the bargaining between the seller and the bidder). Third, the sellersdecision: online, the expected return is $9,867; live, it is ". Thus theseller holds an online auction.

The estimate dispersion has a major eect on the sellers decision.If it is large, the seller may hold a live auction, even given the partici-pation costs. If it is small, it may not pay to sell live.


14/34


Next, we look at sale rates (that is, the percentage of items that areactually sold).

Proposition 5: The sale rate is higher in live auctions than inonline auctions.

In an online auction, since it is a standard ascending auction, theseller sets nontrivial reserve price, thus creating a positive probabilityof no sale. In a live auction, by contrast, since the sellers prot is equalto social welfare net of transaction costs, the seller sets a zero reserveprice and always sells the asset.

We could modify the model to generate the more realistic result thatthe sale rate is less than 100 percent in live auctions by supposing theseller has a positive value for the item unsold. The sellers value might,for example, be given as a draw from the distribution Foff minus some

amount to represent, perhaps, the sellers need for liquidity. The sameintuition applies. Setting a reserve price strictly above the sellers valueof the item unsold pays o in an online auction but not in a live auction.With such a generalization, the sale rate would be less than 100 percentlive, but still lower online.

An alternative eect leading to lower sale rates online is the cost ofresale. Holding another auction is costless online but costly for liveauctions. Online, an unsold item can b e put up for sale again shortlyafterwards. Live, the cost of the preview must be incurred once more.In live auctions, more than two years pass, on average, before a paintingthat failed to meet its reserve nally gets sold (Ashenfelter and Graddy,

2003, Table 3). Thus the seller should be more eager to trade the assetin a live auction than online.

5. Art Auction Data

We now examine data from online and live art auctions in light of ourtheory. We collected data from the Sothebys and eBay websites usingPerl programs. The live-auction data are from Sothebys New Yorksales between June 1 and June 30, 2002, representing 1,890 auctioneditems, with 1,213 successful sales and a total of $23,572,639 raised. Theonline auctions are transactions on the Sothebys site on eBay between

June 26 and July 23, 2002, representing 1,300 auctioned items, with517 successful sales and a total of $68,2845 raised.

Prices are much lower online than live. The highest live-auctionprice in our sample is a George Graham timepiece, at $1,219,500; thehighest online price is a Frank Lloyd Wright copper weed holder (thatis, a vase) at $83,750. The lowest live-auction price is a Cartier watch,at $358; the lowest online price is a drawing by a British artist, Nicole


15/34


Hornby, at $11.50. The mean selling price in the live auctions is justunder $19,000 and in the online auctions just under $1,500.

Histograms of selling prices are given in Charts 1 and 2. Noticehow skewed the distributions are: in both online and live auctionsthe majority of items are quite aordable. This skewedness is typical.Among all artworks auctioned in the United States in 2001, 31 percentsold for less than $1,000, 45 percent for between $1,000 and $10,000,and only 24 percent for more than $10,000.12

The types of items auctioned are compared in Table 1. For the liveauctions, this categorization is based on the auction titles. For theonline auctions, we use eBays categorization.

A higher percentage of successful sales in live auctions than online ispredicted by our theory. The data are in accord with this. Overall, the

sale rate is 64 percent in live auctions and 40 percent in online markets.Table 2 shows the breakdown of sale rates by category. These salerates are roughly consistent with data from elsewhere. For example,in eBays online art auctions in 2000 (the Great Collections auctions,eBays predecessor to its Sothebys joint venture), the sale rate was 48percent (Tully, 2000). At Christies London live auctions in 1995 and1996, the sale rate was in the 70 to 80 percent range for paintings ofvarious kinds, 61 percent for photographs, 88 percent for clocks, and86 percent for jewelry (Ashenfelter and Graddy, 2003).

In auctions run under eBays rules, with a xed end-time, biddersusually submit bids very close to the end of the auction (Roth and

Ockenfels, 2002). In our online-auction data, the mean timing of a bidis 78 percent through the period of auction; the median is at the 96-percent point, and the 75th percentile is at the 99.7-percent point. (SeeChart 3.) In other words, half of the bids, typically, are placed duringthe last 4 percent of the time of the auction, and 25 percent in the last0.3 percent (or roughly the last twenty minutes) of the auction period.Late bidding, therefore, is prevalent not only in the lower-value eBayauctions but also in the art and antiques auctions on Sothebys.com.13

The main result of our theory is that the sellers optimizing decisionto auction online rather than live reects the dispersion of bidder valueestimates as well as, though to a lesser extent, their mean. We run

a simple regression to test this proposition. We use the auctioneerspresale announcement of high and low price estimates to get measures

12According to artprice.com, http://web.artprice.com/EN/AMI/AMI.aspx?id=NTQ3MDEzOTA3MTY513In the early days of Sothebys.com, before the eBay joint venture, the auction

rules a llowed a exible end-time. If a bid arrived within the last ve minutes, theauction was automatically extended for ten minutes; the process repeated untilthere were no more bids. (This was called "popcorn bidding.")


16/34


of the mean and dispersion. We take the mean of these two numbersto be the mean of the estimate distribution, and the dierence between

them to measure the dispersion.14

The highest mean estimate in our live auction data is a Pierre Fred-erich Ingold timepiece, at $375,000 (it was sold); the highest onlineis a Marilyn Monroe wedding gown, at $60,000 (unsold). The lowestmean estimate in our live auctions is a 1995 Cartier watch, at $600(sold); the lowest online is a Lee Tanner photograph of John Coltrane,at $15 (unsold). The highest estimate dispersion in the live auctions isthe same Ingold timepiece, at $250,000; the highest online also is thesame item, the Marilyn Monroe wedding gown, at $20,000. The lowestestimate dispersion in the live auctions is a Fouga wristwatch, at $100(sold); the lowest online is the same Tanner photograph, at $10.

These examples suggest that the pre-auction estimate dispersiondoes indeed reect the degree of valuation uncertainty. A watch orphoto is not unique, so the history of prices of identical or similaritems can be used to assess its value. Moreover, potential buyers canreasonably well judge its value from a written description plus a web-page reproduction. The value of a deceased movie stars wedding dress,by contrast is harder to assess. It is unique, so its future resale value ishighly uncertain. Moreover, it may be hard to judge its quality froma web-page photograph: is it still in a good condition? Seeing it livemay give a bidder signicantly more precise information.

These cases further suggest, as we would expect, that items with a

higher value show a higher dispersion of value estimates. This is thecase in our data. The correlation between the mean of the high andlow estimates and their dierence is 0.9.

A probit regression of the choice of online or live auction regressedon mean and dispersion of estimates is reported in Table 3. Despite thestrong positive correlation between the two independent variables, eachis signicantly dierent from zero. Both the mean and the dispersionof estimates matter. Either a high expected value or a high varianceof value leads to a live rather t han an online sale. Consistent with

14This presumes the auction house reports the presale estimate honestly, not

strategically. For live art auctions, the evidence is mixed. Ashenfelter (1989) ndsestimates are truthful, in that the mean of the high and low estimate is well cor-related with the winning bid. Beggs and Graddy (1997) and Bauwens and Gins-burgh (2000), however, nd systematic underestimation for some kinds of goodsand overestimation for others. Mei and Moses (2002) nd that expensive paintingsestimates are biased upwards. Overall, Ashenfelter and Graddy (2003) conclude,"it is our impression that biases are not quantitatively large when they are preciselyestimated."


17/34


Proposition 4, an item with a given expected value is empirically morelikely to be auctioned online the lower its valuation variance.

6. Simulation

In this section, we examine the eect of transaction costs and infor-mation revelation using a qualitative response model. We assume thatthe sellers utility associated with the choice of the auction is the ex-pected prot

offS and

onS plus an additive error term

off and on. Thedata are the upper and lower estimates x = [x; x]. The parameters weare interested in are the transaction costs cs and cb, and the inspectionparameter k. Let = (c; k). Let Uoff and Uon be the sellers expectedutility: Uoff =

offS (x; ) +

off and Uon = onS (x) + on:

The seller sells the asset in a live auction if Uoff Uon. Thusdening Of f = 1 if the seller sells the asset live,

P(Of f = 1) = P(Uoff Uon) = F(offS (x; x) onS (x;x;))

where F is the distribution function of off on. The log-likelihoodfunction is

log L =nXi=1

Of fi log F(onS (x)

offS (x; ))

+

n

Xi=1

(1 Of fi) log(1 F(onS (x)

off

S (x; ))):

The maximum likelihood estimator^

is dened by @logL@ j=

^

= 0:

We use a simple parametrization. Z=uniform [0:5; 0:5], X = +Z, ui(s; x) =

Pxi=n, and compute the functional form of the discrete

choice model. In an online auction15, since H(w; m) = F(w)n1;

B =1

n(EY1;n EY2;n) =

n(EYZ1;n EY

Z2;n) =

n(n + 1)

S = nB =

(n + 1) :

Note that the sellers expected payo is decreasing in the dispersionand increasing in the number of bidders. Given the smoothness, theconsistency and asymptotic normality of maximum likelihood estimator

15We do not consider reserve prices in this simulation. It i s easy to computeBulow and Klemperer (1996) b ounds.


18/34


is standard (Amemiya, 1985, Section 9.2.2). The number of bidders isdetermined by k=n(n + 1) = cb: For simplicity, we approximate

16 the

solution of this equation by n = (k=cb)0:5

to make the model linear inparameters. We set the number of bidders in the online auction be two.Note 1860/1300=1.43 is the mean number of bidders online. Thus weobtain

onS (x) offS (x; ) = c

0:5b k

0:50:5 + cs

3:

Note that, in this model, we cannot separately estimate k and cb:Now we estimate a probit model of 3190 samples, assuming off

on~N(0; 4002),

of f Coecient Standard Error z P>jzj

c1=2b k1=2 3.011 .7443 4.05 0.000cs 222.2 22.80 9.74 0.000

Table 4: Estimates of Cost and Variance Parameters

A possible value ofc is 90 for k = 0:1: If there are 200 to 300 auctionsin one day for live auctions, then if each bidder bids for ten assets, thetotal participation cost will be around $1000. The threshold valuewhere the bidders rent from selling online is equal to that from sellinglive is $942.30.

7. Business Models

Online auctions have not up to now reached their full potential. Ifour theory is correct, the use of online auctions in the rst few years oftheir existence has been unduly cautious. Sellers have been reluctantto opt for online auctions for high-value items. So far there is notenough evidence that the Internet is right for selling extremely valuableworks. That is still a hard sell, said Sothebys.com executive DavidRedden in 2000.17 According to our model, however, what primarilydetermines whether the seller does better selling live than online is notthe expected price. What matters is the dispersion of valuations, andthe extent to which this dispersion is reduced if the bidders see theitem for themselves. A high-value item can be successfully sold online

if the bidders estimates of that value are tightly bunched.A high dispersion of value tends to go with a high expected value

(in our data set they have a correlation of 0.9). The intuition that

16For reasonably large k, the dierence between the solution of x2 = k andx2

+ x = k is small: for k = 10, the solution is 3.16228 and 2.70156.17Art News, Jan. 2000.


19/34


extremely high value items should not be sold online therefore is typi-cally correctbut it gets the mechanism wrong. High-end items whose

dispersion of value estimates is similar online and live, like the theDeclaration of Independence, are suitable for online sale. It seems thatsellers do not fully recognize thisonline auctions of high-end itemsare after all still quite newand so the limits of online auctions maynot yet have been reached.

Could a seller design an auction to combine the benets of live andonline auctions? An alternative auction form, in between live and onlineauctions, is oered by both Sothebys and Christies: in live auctions,bidders may participate from a distance, by telephone, fax, or internet.The auction-house web sites oer the catalogues for live auctions andallow advance bids, which an auction-house employee will submit dur-

ing the course of the live auction. There is even a provision in some liveauctions for real-time online bidding. Christies oers live streamingvideos, so that people can bid in live auctions from their computers.

From the sellers point of view, however, this in-between alternativeoers none of the saving in transaction costs of true online bidding.Even from a bidders point of view, our analysis implies, this is inferior.Bidders would do better participating either directly in a live auctionor in a true online auction. Distant bidders, who have not incurred thetransaction costs entailed in seeing the piece of art for themselves, arecompeting against the live bidders, who have seen it. Less informedbidders, in common-value auctions in general, are handicapped when

competing against better informed ones, and end up with little or nonet gain (Engelbrecht-Wiggans, Milgrom, and Weber, 1983). Whenthe artwork is of high value, they will be outbid by the better informedlive bidders; when it is of low value, they can win the auction but riskoverpaying. Online bidders in a live auction are in a no-win situation;the theory says they can expect to pay the items full expected value.They can overcome their informational disadvantage by either inspect-ing the item themselves at the presale exhibition or hiring an agent todo the inspection; but this means incurring the same transaction costsas the live bidders. Rather than providing the best of both worlds,mixing online and live bidding combines the disadvantages of both.

8. Conclusion

The sellers choice between online and live auctions involves a trade-o between information generation, favoring a live auction, and trans-action costs, favoring an online auction. Bidders do better online, even


20/34


though they are relatively uninformed about the items true value, be-cause their competitors are equally uninformed and so, from fear of the

winners curse, the bidding competition is less erce. Sellers, also, insome cases do better online because of the saving in transaction costs.

The analysis could be extended in various ways: to incorporate moregeneral assumptions about bidders valuations (aliation rather thancommon value); to derive endogenous participation costs in the pres-ence of bidder asymmetry; and to estimate participation costs andinformation revelation rates.


21/34


9. Appendix: Proofs of the Propositions

Proof of Lemma 1: First, compute the expected prot of thebidder and the seller in a standard ascending auction of the generalsymmetric model with n bidders and a signal distribution f. In an as-cending auction, there exists a symmetric monotone equilibrium whereeach bidder drops out at a price equal to the expected value of the assetgiven the signal of bidders who have already dropped out and assumingthat the bidders remaining in the auction have the same signal withthe bidder (Milgrom and Weber, 1982a). Compute an expected payoB(xi) of a bidder i with a signal xi. By symmetry, B is independentof the identity of the bidder. Suppose bidder i with signal xi bids as ifher signal is yi. Let the expected prot be B(xi; yi). Then

B(xi; yi) =

ZXi

(v(xi; xi) v(y1;i; xi))f(xijy1;i yi)dxi

where v(x) = E[sjX = x].and y1;i is the highest signal among biddersother than i. This is because bidder i wins if and only if bidder isclaimed type yi is higher than the other bidders. If bidder i wins, thevalue is E[sjX = x]:

dB(xi)dxi

= dB(xi; xi)dxi

=Z

@B(w; y)@w

jy=wdw:

This is because, since v and f are smooth and bounded, we can applythe envelope theorem. The bidder with the lowest type never wins, soher expected prot is zero. Then,

B(xi) =

Zxi1

ZXi

@vi(w; xi)

@wf(xijy1;i w)dxidw

=Zxi1

H(w; n)dw

(where H(w; n) =

ZXi

@vi(w; xi)

@wf(xijy1;i w)dxi:)

Then the bidders ex ante payo is:


22/34


B = Z Zxi

1

H(w; n)dwf(xi)dxi

= f

Zxi1

H(w; n)dwF(xi) g+11

ZHi(xi; n)F(xi)dxi

=

ZH(w; n)dw

ZHi(xi; n)F(xi)dxi

=

ZH(xi; n)dxi

ZHi(xi; n)F(xi)dxi

=

ZH(xi; n)(1 F(xi))dxi:

The rst line is by taking the expectation of B

(xi) and changing the

order of integration. The second line is by integration by parts. Thethird line is by integration by parts using the formula

Rab = [ab]

Ra0b

with a =R

H(w)dxi and b = F(x): The sellers ex ante prot is thedierence between the items expected value and the bidders expectedprots:

S = Es nBSecond, we compute expected prots in an online auction without a

reserve price. From the previous formula, but with N bidders,

onB = Z Hon(xi; N)(1 F

on(xi))dxi:

onS = Es N

ZHon(xi; N)(1 F

on(xi))dxi:

The number of bidders in the live auction is determined by the zero-prot condition (with bidders entering until prots are competed away,ignoring the integer constraint):

offB =

ZHoff(xi; n

of f)(1 Fof f(xi))dxi = cb

The sellers ex ante expected live-auction prot is

offS = Es noffcb cs

= Es noffZ

Hoff(xi; noff)(1 Foff(xi))dxi cs:

Proof of Proposition 1: First, we claim that it is suce to showthat bidders expected prot is lower in an auction with 1 than onewith 2. This is because the sellers expected prot is the dierence


23/34


between the value of the asset and the bidders expected prot. (Bythe assumption of absolute common value, the value of the asset is

independent of .)Second, we compute the lower and upper bound of the bidders ex-

pected prot. Recall from Lemma 1, the bidders expected prot is

B =

ZH(xi; n)(1 F(xi))dxi:

Thus, from the denition of H and the signal-sensitivity condition(3.1), B is bounded:

m

ZF(xi)

n1(1 F(xi))dxi B M

ZF(xi)

n1(1 F(xi))dxi:

The distributions and densities of the order statistics are (David, 1981):

F1;n(x) = F(x)n

F2;n(x) = F(x)n + nF(x)n1(1 F(x))

f2;n(x) = n(n 1)(1 F(x))F(x)n1f(x)

and therefore a bidders expected prot can be written as:

ZF(xi)

n1(1 F(xi))dxi = (1=n)

Zn(F(xi)

n1 F(xi)n)dxi

= (1=n)Z

(F2;n(xi) F1;n(xi))dxi

= (1=n)

Zxi(f1;n(xi) f(x2;n(xi))dxi

=1

n(EY1;n EY2;n)

=

n(EZ1;n EZ2;n):

where the third equality uses integration by parts, the fouth is bydenition, and the last comes from the properties of mean-dispersionfamilies (where Z is the base random variable as dened above).

Third, we show that the bidders prot is lower in auctions withdispersion parameter 1: The upper bound of the bidders prot froman auction with dispersion parameter 1 is 1M=n(EZ1;nEZ2;n): Thelower bound is 2m=n(EZ1;n EZ2:n). Thus the bidders prot froman auction with 1 is lower if

1M=n(EZ1;n EZ2;n) 2m=n(EZ1;n EZ2:n)


24/34


which implies

(9.1) 1 < (m=M)2:

Proof of Proposition 2: First, we claim B is monotone decreasingin n. The sellers prot is the highest marginal revenue among bidders.Then the addition of one bidder weakly increases the sellers prot.This implies that the total surplus for the bidders decreases. Thus thebidders ex ante prot decreases.

Second, we claim that n is monotone increasing in :The numberof bidders n is determined by the zero prot condition: B(n; ) = c.From the previous discussion, B(n; ) is monotone increasing in andmonotone decreasing in n. Thus n is monotone increasing in :

Third, we claim that the sellers expected prot is monotone de-creasing in . From Lemma 1, the sellers expected payo is the valueof the asset minus the total payment of entry costs. By the absolutecommon value assumption, the value of the asset is constant and thetotal payment of entry cost is monotone decreasing in .

Proof of Proposition 3: From Lemma 1, and given that m = Mas a result of the mean-common-value assumption, a bidders expectedprot is a constant multiplied by the dierence in the order statistics,(EY1;n EY2;n), which is independent of the mean parameter .

Proof of Proposition 4: First, we compute an upper bound ofthe sellers prot from the online auction. By Bulow and Klemperer

(1996), the sellers expected prot from an auction with n bidders withan optimally dynamically set reserve price is less than the expectedprot from n + 1 bidders without a reserve price. Thus,

onS = V (N + 1)

ZHon(xi; N + 1)(1 F

on(xi))dxi:

V (N + 1)m

ZFon(xi)

N(1 Fon(xi))dxi

= V m(EYon1;N+1 EYon2;N+1)

= V m(EYZ1;N+1 EYZ2;N+1):

Second, we compute a lower bound on the sellers prot in a liveauction with N bidders. By the signal-sensitivity condition (3.1), thereexists M such that @v(x)=@xi M. Thus:

offS V kM(EYZ1;N EY

Z2;n) (N+ 1)cb cs:


25/34


Third, assuming there are N bidders in the live auction, we computek0 and 0 such that for k k0 and 0 ; a bidders rent is smaller in

the live auction. This holds ifV k M(EYZ1;N EY

Z2;n) (N + 1)cb cs V m(EY

Z1;N+1

EYZ2;N+1)

kM(EYZ1;N EYZ2;n) + (N + 1)cb + cs m(EY

Z1;N+1 EY

Z2;N+1)

(m(EYZ1;N+1 EYZ2;N+1) kM(EY

Z1;N EY

Z2;n)) (N+ 1)cb + cs

Thus for

k m(EYZ1;N+1 EYZ2;N+1)=M(EY

Z1;N EY

Z2;n)

if

[(N + 1)cb + cs]=[m(EYZ1;N+1 EY

Z2;N+1) kM(EY

Z1;N EY

Z2;n)];

the sellers payo is higher in the live auction.Fourth, we compute the conditions under which at least N bidders

enter the live auction. This occurs if each bidders expected payo withN bidders is higher than the entry cost:

offB m RFon(xi)N(1Fon(xi))dxi =

mN

k(EYZ1;N+1 EYZ2;N+1)

cb:That is,

Ncb

mk(EYZ1;N+1 EYZ2;N+1)

Fifth, we nd when both conditions are satised. For each k thatsatises k m(EYZ1;N+1 EY

Z2;N+1)=M(EY

Z1;N EY

Z2;n); there exists

max[Ncb

mk(EYZ1;N+1 EYZ2;N+1)

;(N + 1)cb + cs

m(EYZ1;N+1 EYZ2;N+1) kM(EY

Z1;N EY

Z2;n)

]

where the sellers prot is higher in the live auction. In addition, if ishigh enough that expected value of the allocation V() is higher thancbN + cs;the seller earns a positive prot.

Proof of Proposition 5: In an online auction, by Lemma 2 ofBulow and Klemperer (1996), the sellers reserve price is the price thatmakes marginal revenue equal to zero. Thus, if the bidder with thehighest marginal revenue is less than zero, the seller does not trade.


26/34


Second, we compute the reserve price in a live auction. If the zeroprot condition is binding, the sellers expected prot is equal to the

expected social surplus, thus the seller, who has the value zero for theasset, does not trade. If the zero prot condition is not binding, theneach bidders marginal revenue is strictly higher than that in an onlineauction, thus the probability of exclusion is strictly lower.


27/34


10. Tables

Table 1: Number of auctions by category

Live OnlineJewelry 635 73Paintings and sculpture 621 113Clocks and watches 562 41Antiques and furniture 286 233Books and prints 283 37Silver and ceramics 0 259Photos, stamps, coins 0 272Collectibles 0 356

Table 2: Sale ratesLive Online

Paintings and prints 83% 44%Watches and clocks 82% 63%

Jewelry 54% n.a.Furniture n.a. 66%

Coins and silver n.a. 80%Books 57% n.a.

(n.a. means no data available)

Table 3: Regression of online vs. live

O Co ecient Standard error z P>jzjEstAvg 0.0000686 0.0000135 5.09 0EstDi 0.0000336 0.00000461 7.30 0Const -0.493068 0.0325403 -15.15 0

Obs 3180LR 1207.75

Prob>chi2 0PseudoR2 0.2801


28/34

Figure 1: A Painting Offered on Sothebys.com

Estimate: $25,000 to $35,000

Seller: gsauer (353) View sellers feedback

Description:An oil on canvas painting by William Clapp in a custom wood frame withgold leaf. An old label verso states: "Guaranteed Authentic, William H. Clapp, C-691142,Lakey Gallery, Carmel". [Lakey Gallery handled the artist's works in the 1960s]. It shouldbe noted that this painting is one of the largest he produced.Notes: William Henry Clapp [1879-1954] was born in Montreal, Canada and moved toOakland, California with his family in 1885. He returned with his family to Montreal in1900 to study at the Montreal Art Association School. In 1904, he went to Paris andstudied at the Ecole des Beaux Arts and Academie Julian. He was strongly influence bythe French Impressionists and Pointillists at the time. He returned to Montreal in1906and in 1917 moved with his family to Piedmont, California where he became the directorof the Oakland Art Gallery from 1918 to 1949. At this time, be became an aggressive

force for modernism and experimentation. As a member of the Society of Six he helpedcreate a modernist style that was uniquely Californian and a rebellion against thatespoused by the tonalist styles of Arthur Mathews and William Keith. His works are innumerous institutions.Condition: In fine condition.Measurements: 30"x36"; frame 40"x46"


29/34

Figure 2: A Clock Offered on Sothebys.com

Estimate: $2,000 to $3,000

Seller: empiregalleryla (36) View sellers feedback

Description: Charles X Alabaster Portico Clock. 19th century. The architecturalcase centered by a circular gilt bronze dial inscribed "Vollabys Dany' ADunkerque", and displaying a rosette pendulum on bun feet.Notes: Mid 19th CenturyCondition: Small scratch and chips repolished, overall very good condition.Measurements: w 11", h 20", d 6 1/2"

Source of Figures 1 and 2: http://sothebys.ebay.com, January 14, 2003
http://cgi2.sothebys.ebay.com/aw-cgi/eBayISAPI.dll?ViewFeedback&userid=gsauer&ssPageName=L2http://cgi2.sothebys.ebay.com/aw-cgi/eBayISAPI.dll?ViewFeedback&userid=gsauer&ssPageName=L2


30/34

References

Amemiya, Takeshi (1985),Advanced Econometrics, Harvard University Press

Ashenfelter, Orley (1989), How Auctions Work for Wine and Art,Journal of

Economic Perspectives 3(3), 23-36.

Ashenfelter, Orley and Kathryn Graddy, Auctions and the Price of Art,Journal ofEconomic Literature, to appear, 2003.

Athey, Susan and Jonathan Levin (2000), The Value of Information in MonotoneDecision Problems, unpublished, Stanford.

Bajari, Patrick and Ali Hortacsu (2002a), The Winner's Curse, Reserve Prices and

Endogenous Entry: Empirical Insights from eBay Auctions, forthcoming,Rand Journalof Economics.

Bajari, Patrick and Ali Hortacsu (2002b), Cyberspace Auctions and Pricing Issues: A

Review of Empirical Findings, forthcoming in theNew Economy Handbook.Baron, David P., Private Ordering on the Internet: The eBay Community of Traders,

unpublished, GSB Stanford, 2001

Baye, Michael, John Morgan, and Patrick Scholten (2001), Price Dispersion in the Smalland Large: Evidence from an Internet Price Comparison Site, unpublished, UCB Haas.

Beggs, A. and K. Graddy (1997), Declining Values and the Afternoon Effect: Evidence

from Art Auctions,Rand Journal of Economics, 28, 544-65Brown, Jeffrey R., and Goolsbee, Austan, Does the Internet Make Markets More

Competitive? Evidence from the Life Insurance Industry,Journal of Political Economy

110 (3), 2002, 481-507Brynjolfsson, E., and Smith, M. (2000), Frictionless Commerce? A Comparison of

Internet and Conventional Retailers,Management Science, Vol. 46, No. 4

Bulow, Jeremy and Paul Klemperer (1996), Auctions versus Negotiations,AmericanEconomic Review 86(1), 180-93.

Bulow, Jeremy and John Roberts (1988), The Simple Economics of Optimal Auctions,Journal of Political Economy.

David, Herbert (1981), Order Statistics, Wiley.

Duret-Robert, Franois, The Verdict of the Sale-Room, in Maurice Srullaz, ed., ThePhaidon Encyclopedia of Impressionism, Oxford, Phaidon Press, 1978

Ellison, Glenn, and Ellison, Sara F., Search, Obfuscation, and Price Elasticities on the

Internet, unpublished, Department of Economics, MIT, June 2001Engelbrecht-Wiggans, Richard, Paul R. Milgrom, and Robert J. Weber, Competitive

Bidding and Proprietary Information,Journal of Mathematical Economics 11, 1983,

161-69.Hallowell, Roger (2001), Sothebys.com, Harvard Business School case 9-800-387.

Hausch, D.B., and L. Li (1993), A Common Value Auction Model with Endogenous

Entry and Information Acquisition,Economic Theory 3(2), 315-334.Hildesley, C. Hugh, Sotheby's Guide to Buying and Selling at Auction, New York,

Norton, 1984.

Koppius, Otto, Van Heck, Eric, and Wolters, Matthijs, The Importance of Product

Representation Online: Empirical Results and Implications for Electronic Markets,unpublished, Erasmus University, Rotterdam, 2002


31/34

Levin, Dan and James Smith (1994), Equilibrium in Auctions with Entry, American

Economic Review 585-599.

Lucking-Reiley, David, Doug Bryan, Naghi Prasad, and Daniel Reeves, Pennies from

eBay: the Determinants of Price in Online Auctions, unpublished, University ofArizona, 2000

McAfee, R. Preston, and McMillan, John, "Auctions with Entry,"Economics Letters 23,1987, 343-48.

McAfee, Preston, Daniel Quan, and Daniel Vincent, How to Set Minimum AcceptableBids, with An Application to Real Estate Auctions, Journal of Industrial Economics 50

(4), December 2002

McAfee, Preston, and Daniel Vincent, Updating the Reserve Price in Common ValueAuctions, American Economic Review Papers and Proceedings, 82, May 1992, 512-8

McDonald, Cynthia G. and V. Carlos Slawson, Jr., Reputation in an Internet Auction

Model, unpublished, University of Missouri-Columbia, 2000McMillan, John,Reinventing the Bazaar: A Natural History of Markets, New York,

Norton, 2002

Mei, J. and Moses, M. (2002) Are Investors Credulous? Some Preliminary Evidencefrom Art Auctions, unpublished, Stern School of Business, New York University

Milgrom, Paul and Robert J. Weber (1982a), A Theory of Auctions and Competitive

Bidding,Econometrica 50(5), 1089-1122.Milgrom, Paul and Robert J. Weber (1982b), The Value of Information in a Sealed-Bid

Auction,Journal of Mathematical Economics 10, 105-114.

Pesando, James E., Art as Investment: The Market for Modern Prints,AmericanEconomic Review 83, December 1993, 1075-89

Resnick, Paul and Richard Zeckhauser, Trust Among Strangers in Internet

Transactions: Empirical Analysis of eBays Reputation System, unpublished,University of Michigan, 2001

Roth, Alvin E. and Axel Ockenfels (2002), Last-Minute Bidding and the Rules forEnding Second-Price Auctions: Evidence from eBay and Amazon Auctions on the

Internet,American Economic Review, forthcoming

Sothebys (1988), Contemporary Prints: New York, November 5, 1988, Sotheby's.Tully, Judd (2000), Upmarket eBay and Cyber Sothebys, www.cigaraficionado.com

Turban, Efriam, Auctions and Bidding on the Internet: An Assessment, Electronic

Markets 7 (4), 1997, 7-11Watson, Peter,From Manet to Matisse: The Rise of the Modern Art Market, New York,

Random House, 1992

Wilson, Robert B., A Bidding Model of Perfect Competition,Review of EconomicStudies 44, 1977, 511-518

Ye, Lixin (2001), Optimal Auctions with Endogenous Entry, unpublished


32/34


33/34


34/34

Documents

The Value of a Good Repution Online (Application to Art Auctions)