BEAUTY CONTEST IN FINANCIAL MARKETS: AN EXPERIMENT WITH STOCK MARKET PROFESSIONALS Jukka Ilomäk * School of Management Unversty of Tampere Ths verson: January 30, 2013 Abstract We study the behavor of stock market professonals wth a short horzon assumpton n the expermental settngs. A novelty n the expermental desgn s the use of the endogenous equlbrum prce system wth a fundamental value that vares n tme. Informed subjects are told the fundamental value of the next perod. The results suggest that there s a statstcally sgnfcant sentment-based beauty contest effect on the forecastng behavor of nformed and unnformed market professonals. In both experments, the equlbrum prce drfts away from the fundamental value. I apprecate helpful comments from John Lst, Jar Vanomäk and Juha Junttla, and I gratefully acknowledge fnancal support from Nasdaq OMX Nordc Foundaton and the Fnnsh Foundaton for Share Promoton. I also thank Reeta Hemmnk and Ann Harmaala who helped to run the experment. 1
1. INTRODUCTION Keynes [1936] noted that a ratonal nvestor n the fnancal markets s nterested n the belefs of other nvestors about the future returns of the asset, and the belefs of other nvestors about that belef, and so on. In other words, nvestors care about hgher order belefs, and they care about belefs other than ther own. Keynes calls ths phenomenon a beauty contest n the fnancal markets. The hgher order belefs can be consdered as sentment-based nformaton f the nformaton comes from past returns. Accordng to the effcent market hypothess (EMH), sentment-based hgher order belefs do not have any nfluence n the asset market equlbrum. Accordng to the EMH, even n an nvestng perod wth a very short horzon a ratonal nvestor should behave as an agent wth a long horzon and an nfnte lfe. That s, she should speculate on long horzon fundamentals. Ths can be proved by backward nducton [please see, for example, Froot, Scharfsten and Sten 1992]. Allen, Morrs and Shn [2006] argue that n a one-asset economy, the equlbrum prce of the asset may dffer from ts constant fundamental value under nosy expectatons. They nterpret ths as a beauty contest phenomenon and as the effect of the hgher order belefs. Bacchetta and van Wncoop [2008] use a ratonal expectatons model wth nfnte tme perods, where the short-lved nvestors have heterogeneous nformaton about the asset and the fundamental value of the asset vares over tme. They show that the hgher order belefs can pull the equlbrum prce away from the fundamental value of the asset. 2
In real-lfe asset markets, t may be dffcult to do an exact separaton of the fundamental and the sentment-based nformaton that moves stock markets. In addton, the prvate nformaton of nvestors s unobservable. In expermental settngs, we can control the nformaton flows that affect nvestors decsons. In ths expermental study, we want to study the followng: Can we fnd evdence of the sentment-based beauty contest effect n an expermental settng where the subjects are stock market professonals? In our experment, we use artfcal data from the fnancal markets (random walk wthout a drft) as the fundamental value of the asset, and ask stock market professonals to forecast the equlbrum prce at the next step ahead wth the avalable nformaton. Thus, we have a short horzon market experment where the fundamental value vares over tme. We splt the subjects nto two groups where the rato of nformed to unnformed subjects s dfferent, and splt the groups nto two pools (nformed and unnformed subjects). One subject pool n each group receves consstently relable prvate nformaton n addton to publc nformaton. That s, they receve relable prvate nformaton about the fundamental value of the asset that follows the random walk (n advance of the next step ahead), and nformaton about the past prces. The subjects n the other (unnformed) pool n the group are purely nose traders whose only nformaton s the past prces of the asset n queston. The EMH predcts that a trader whose only nformaton s the past prces should forecast the prces accordng to a martngale model, and that an nformed trader should use her relable prvate nformaton consstently and gnore past prce changes. 3
In ths expermental desgn, we can detect a sentment-based beauty contest phenomenon n the forecasted prces of the market professonals. Ths s the observable dfference between the subject s last forecasted return and the last equlbrum return, because the return dfference s the subject s forecasted error n the prevous prce change. The EMH mples that the past returns or past forecasts of the asset (n our settng) should not explan the changes n the forecasts. The results suggest that n short horzon markets a market professonal tends to care about belefs other than her own. She cares about what other market partcpants have been estmatng. Ths s the sentment-based beauty contest effect on the fnancal market n ths experment. It does not matter f she has prvate fundamental nformaton about the true value of the asset or not. Accordng to the results, an nformed market professonal utlzes, n addton to the beauty contest phenomenon, her prvate nformaton about the true fundamental value of the asset only f she s n a majorty n the market. The results are consstent wth the theoretcal hgher order belefs n the fnancal market lterature. In both experments, the equlbrum prce drfts away from the fundamental value. The results ndcate that the sentment-based beauty contest effect occurs, at a 0.001 % sgnfcance level, n three out of the four pools. Secton 2 outlnes the related lterature. Secton 3 presents the expermental desgn, Secton 4 presents an analyss of the results and Secton 5 concludes. 4
2. RELATED LITERATURE 2.1 THEORETICAL BEAUTY CONTESTS IN FINANCIAL MARKETS LITERATURE The theoretcal studes of Froot, Scharfsten and Sten [1992], Abreu and Brunnermeer [2003], Allen, Morrs, and Shn [2006], Cpran and Guarno [2008], Bacchetta and van Wncoop [2008] and Banerjee, Kanel, and Kremer [2009], among others, can be consdered as beauty contest papers. That s, ther common feature s that the nvestors try to estmate what other nvestors are dong and ths drves equlbrum prces away from the fundamental values. Froot, Scharfsten and Sten [1992] show that rsk neutral nformed and nose traders wth short horzons can drve the equlbrum prce away from the fundamental value, f the majorty of the traders n acton are techncal analyss traders and the traders assume that they can close ther poston before the fundamental value of the asset s publcly announced. Abreu and Brunnermeer [2003] show that, as nformed nvestors are unable to coordnate ther sales, the equlbrum prce can drft above the fundamental value for long perods. Abreu and Brunnermeer have an economy wth one rsky asset and an nfnte tme horzon, but they assume that the bubble wll burst n at least a fnte tme for some exogenous reason and that the equlbrum prce then goes back to the fundamental value. The nvestors n ther model are rsk neutral, but the nablty of nformed nvestors to act n a coordnated way, and the fnte tme horzon of the 5
bubble, prevent the nformed nvestors from takng nfnte short postons aganst the bubble. Allen, Morrs, and Shn [2006] show that n a fnte economy wth one rsky asset, where ratonal short-lved nvestors are rsk-averse, the constant fundamental value and the equlbrum prce may drft apart for long perods, because sequental equlbrum prces n a one-asset economy create a momentum effect. Cpran and Guarno [2008] study sequental tradng n an nfnte economy wth two rsky assets. In ther model, nvestors trade n an exogenously determned sequental order and nteract alone wth a market maker. These authors show that herdng and nformatonal cascades can occur n one market and be transmtted from one market to another. They defne herdng as a phenomenon where nvestors choose the same acton, and an nformatonal cascade as a phenomenon where there s a complete blockage of fundamental nformaton. In other words, durng an nformatonal cascade trades do not convey any nformaton about the true value of an asset. Bacchetta and van Wncoop [2008] have a ratonal expectatons model wth nfnte tme perods, where the short-lved rsk-averse nvestors have heterogeneous prvate nformaton about the fundamental value of the asset that vares wth tme. In ther model, the fundamental value and the equlbrum prce drft apart, because the nvestors are rsk-averse and have heterogeneous nformaton about the fundamental value of the asset. These authors call the dfference between the equlbrum prce and the fundamental value of the asset at tme t as the hgher order wedge, whch s caused by hgher order belefs. 6
Banerjee, Kanel, and Kremer [2009] have a fnte economy wth one rsky asset, where ratonal nvestors are rsk-averse. These authors argue that n ther settng prce drft only occurs f the nvestors have a hgher order dfference of opnons. Hence n ther model, the equlbrum prce and the fundamental value of the asset may drft apart only f the nvestors have dfferent estmatons about the common estmaton of the wedge. 2.2 EXPERIMENTAL FINANCE LITERATURE The expermental studes about fnancal markets nclude Hagh and Lst [2005], Drehmann, Oechssler and Roder [2005] Alevy, Hagh and Lst [2007] Cpran and Guarno [2005], Cpran and Guarno [2009], Hommes et al. [2005], Hagh and Lst [2010], Ilomäk [2012] among others. Hagh and Lst [2005] carred out a lottery bet experment wth market professonals and students, and tested whether the subjects were affected by the myopc loss averson (MLA) phenomenon. They found clear evdence of MLA, and concluded that market professonals exhbt MLA to a greater extent than do students. Drehmann, Oechssler and Roder [2005] performed a large nternet experment wth general professonals and students to test herdng behavor n the fnancal market settng. In ther experment, the subjects had to decde, sequentally n some exogenous order, whether to nvest n one of the two assets, A or B. No evdence of herdng was found. 7
Alevy, Hagh and Lst [2007] nvestgated nformatonal cascades wth fnancal market professonals and wth students. They follow the procedure of Anderson and Holt [1997] n whch t s possble to control nformaton flows by drawng balls from an unobservable urn, and to determne whether the subjects tend to follow prevous decsons only when t s ratonal to do so. They concluded that market professonals are less Bayesan than are students, but that they base ther decsons on the qualty of others decsons to a greater extent than do students. Cpran and Guarno [2005] studed herd behavor n a laboratory fnancal market. In ther experment, undergraduate students receved relable prvate nformaton on the fundamental value of an asset and traded t sequentally n some exogenous order wth a market maker. The authors found that subjects seldom herded but more frequently preferred to gnore ther prvate nformaton and abstan from tradng. Cpran and Guarno [2009] re-executed ther 2005 experment wth fnancal market professonals. In addton, they added a new element to ther desgn so that herdng becomes optmal because of event uncertanty, and decsons are collected from all partcpants at every tme step. In the 2005 paper, they had allowed the subjects reveal ther decsons only once, at the tme of ther turn to trade. The authors found that herdng behavor exsts but not to such an extent as would make t optmal. Hagh and Lst [2010] tested the behavor of market professonals and students n an expermental optons model settng. They found that the behavor of both groups was more consstent wth the predctons of the optons model than wth those of the 8
classcal nvestment model. However, students tend to be more senstve to payoff changes than are market professonals. Experments wth stock market professonals nclude Hagh and Lst [2005], Alevy, Hagh and Lst [2007], Cpran and Guarno [2009], Hagh and Lst [2010] and Ilomäk [2012] among others. Ilomäk [2012] studed the behavor of stock market professonals n an experment wth real fnancal market data and real prvate nformaton. In the experment he utlzes actual data from the fnancal markets and real prvate nformaton, and puts the stock market professonal n a hedge fund manager role where she has to take a long or short poston on the asset sequentally. He splts the subjects randomly nto two groups (nformed and unnformed traders), and he assumes that the subjects are homogenous wthn each pool because they are all experenced stock market professonals. One subject pool receves consstently relable prvate nformaton and past returns, and the other pool contans purely nose traders where the only nformaton s past returns from the fnancal markets. Ilomäk found that the unnformed subjects utlze the U.S. returns to forecast the next return of the small country. The nformed subjects use both the U.S. returns and ther prvate nformaton. In addton, they gve the U.S. returns the most nfluence. He found the affect heurstc n the forecasts. 2.3 BEAUTY CONTEST EXPERIMENTS Nagel [1995] proposes a laboratory beauty contest where the subjects try to guess a real number between 0 and 100. The wnner s the subject(s) whose number s the closest to a number x*, whch Nagel defned as p tmes the average of the guesses of 9
the subjects, where 0 < p < 1. Kocher and Sutter [2005] studed the beauty contest phenomenon n the laboratory expermental settng by a smlar procedure and, n addton, they compared the behavor of small groups n the beauty contest. Kocher and Sutter found evdence of a beauty contest phenomenon n ther laboratory experments. Hommes et al. [2005] propose a beauty contest experment n the stock market settng, where the task of the subjects (undergraduate students) s to forecast the prce level of asset X at the next step ahead. The only nformaton for the subjects s the past prces and the constant fundamental value. The subjects are not aware that the equlbrum prce that s acheved s an average of the subjects forecasts at step t. Thus, they have one treatment n the experment. The authors found clear evdence of the beauty contest phenomenon n ther settng. 3. EXPERIMENTAL DESIGN Our man dea s to follow the procedure of Ilomäk [2012], and make the equlbrum prces endogenous as suggested by Hommes et al. [2005], to detect a sentment-based beauty contest phenomenon. In contrast to Hommes et al. we use stock market professonals, we have treatments for nformed and unnformed subjects and the fundamental value vares wth tme n our experment. 10
3.1 SUBJECTS The followng experment belongs to the framed feld experment category and was carred out wth stock market professonals from Scandnavan fnancal nsttutons. The subjects (twenty partcpants) are experenced (wth a mnmum of fve years experence) professonal traders and nvestment bankers wth a master s degree n economcs or fnance. There are sxteen men and four women n the expermental groups. In order to acheve permsson to run ths experment we were forced to apply a strct anonymty polcy concernng the partcpants. Thus, we have only general nformaton about the characterstcs of the subjects. 3.2 TASK We splt the subjects nto two groups (nne and eleven subjects). Then we randomly splt these groups nto two pools (nformed and unnformed professonals). In the frst group we have fve (56 %) nformed and four (44 %) unnformed subjects. In the second group we have fve (45 %) nformed and sx (55 %) unnformed subjects. Thus we execute the experment twce, wth the prvate nformaton beng gven or not gven, and wth the dfferent partcpants. The prvate nformaton s the same for all nformed subjects, but they are not explctly told ths fact. The subjects are told that ther task s to forecast the prce of the asset X at the next step ahead. There are 24 forecastng perods. The unnformed subjects only nformaton (common to all) s the prevous equlbrum prce of the asset X, whch they can wrte down as the experment goes on. 11
The realzed prce s an average of the subjects forecasts usng the avalable nformaton at step t. (The subjects are not explctly told ths fact). The subjects can use ther computers to calculate returns and execute analyses. At the begnnng, all subjects are told that the last prce of the asset X s 40. Then after the subjects have made ther predctons about the prce for perod one, they receve nformaton about the realzed prce of the asset X at perod one. That s, they receve the average of ther own forecasts, but the fact that t s actually ther average s not revealed to the subjects. Thus, the subjects form an artfcal fnancal market. The same procedure s repeated as the experment goes on. Every subject knows hs or her own forecastng accuracy after each perod (24 perods). 3.3 PRIVATE INFORMATION The prvate nformaton for the nformed nvestors s the followng. We reveal that the fundamental value of the asset X s perfectly correlated wth an unspecfed commodty prce. We tell the nformed subjects that the prce of asset X s equal to ts fundamental value at perod zero (40 ). We reveal the prce change of the unspecfed commodty prce at the next step ahead. The fundamental nformaton s the same for all nformed subjects. The fact that t s actually a random walk s not revealed. 12
3.4 PAYOFFS There are 24 tme steps to forecast. The subject earns the observed return f she correctly predcts the sgn of the return, and earns twce the observed return f she correctly predcts the return (n whole numbers). The subject loses the realzed return f she wrongly predcts the sgn of the return. Communcaton between the subjects s prohbted, and the subjects are seated so that no ndvdual can observe another ndvdual s decson and payoffs durng the experment. Thus, the partcpants cannot get any feedback on how or what other partcpants are dong durng the experment. Naturally, the subjects wll notce ther own performance as the experment goes on. However, after the experment nobody has to redeem her possbly negatve cumulatve payoffs. The methodology of the experment can be found n the Appendx. 4. THE EXPERIMENT We ran the experment n two major Nordc fnancal nsttutons, between March 16 and Aprl 26, 2012, n two separate sessons. The partcpants were experenced stock market professonals (twenty partcpants who were traders or nvestment bankers). The sessons started wth the followng wrtten nstructons gven to all subjects: Your task s to forecast the next step ahead equlbrum prce for the asset X as a whole number. At perod zero, the equlbrum prce s 40. There are 24 tme steps to forecast. After all partcpants have made ther forecasts for a perod, the answer papers wll be collected and after that the realzed prce 13
wll be revealed for that perod. Communcaton between subjects s prohbted. You can use your computer freely to calculate returns and execute analyses as the experment goes on. Payoffs: If you forecast the return correctly, your payoff wll be twce the realzed return. There are no short sellng costs. If you forecast the sgn of the return correctly, you earn the realzed return. If you forecast the sgn of the return wrongly, you lose the realzed return. However, after the experment nobody has to redeem her possbly negatve cumulatve payoffs. In addton, the nformed nvestor pool got the followng nformaton. Prvate nformaton: The fundamental value of the asset X s perfectly correlated wth an unspecfed commodty prce. The prce of the asset X s equal to ts fundamental value at perod zero. You have relable prvate nformaton that the expected prce change of the commodty prce for the forecastng perod s. 4.1 ANALYSIS OF THE RESULTS 4.1.1 THEORETICAL PREDICTIONS Accordng to the EMH, the nformed nvestor should use her relable prvate nformaton consstently and gnore past prce changes and past forecasted prce changes. The unnformed nvestor should forecast accordng to the martngale model: 14
E X t t X t 1 (1) where θ t s an nformaton set and X s a stock prce. The martngale property mples only that the expected values of future prce changes wll be ndependent of the values of past prce changes. It means that the next prce change s equally lkely to be postve or negatve, and that all forecastng rules for the next perod whch are based on the hstorcal returns of stock prces should be useless. Thus, t mples that the best forecast of tomorrow s prce s today s prce. Accordng to the theoretcal fnancal markets beauty contest lterature, wth short nvestng perods ratonal nvestors care about hgher order belefs. That s, they try to estmate what other nvestors are estmatng. In our settng, the sentment-based beauty contest effect s the observable (by the subject) dfference between the last forecasted return of the subject and the last equlbrum return, because the return dfference s her forecasted error for the prevous prce change. Thus, ths s the latest nformaton for the subject about the dfference n opnon between the market and herself. Ths s the effect of the past returns and we can measure ths statstcally. In our expermental settngs, we have four pools, and the experenced subjects get homogeneous nformaton wthn the pools. In the settngs, an nformed nvestor knows the next perod return and prce of the fundamental value of the asset X, the last equlbrum prce and last return of the asset X and, naturally, her last forecasted prce and her last forecasted return. An unnformed nvestor knows the last equlbrum prce and last return of the asset X and her last forecasted prce and her last forecasted return. Thus, n the settng, we must assume n both cases (nformed 15
and unnformed) that the sentment-based beauty contest effect s the dfference between the trader s last forecasted return and the last equlbrum return, because the return dfference s her forecasted error for the prevous prce change. The EMH predcts that the prevous prce change does not affect the next prce change. Thus, t would be evdence for the sentment-based beauty contest phenomenon n the experment f we found ths varable statstcally sgnfcant n the experment. Defnton 1. The sentment-based beauty contest effect n ths experment = the dfference between the last forecasted return of the subject and the last equlbrum return. 4.1.2 DESCRIPTIVE STATISTICS We have two groups (nne and eleven subjects) and two pools n each group. We calculate the aggregate forecasts for the groups for every tme step. Ths makes 24 aggregated forecasts for both pools and the two groups. In addton, we have the equlbrum prce for the groups (average forecast of the prce) at every tme step and we have the fundamental value of the asset X, whch s the same for both groups. Fgure (1) shows the actual tmes seres of the frst group (Group I), where 56 % are nformed subjects (beng gven the same prvate nformaton) and 44 % are unnformed subjects (whose only nformaton s the hstory of equlbrum prces). 16
Fgure (1). Forecasts of nformed subjects (56 %), unnformed subjects (44 %), equlbrum prces and fundamental value prces n Group I. In the second group (Group II) there are fve nformed subjects (45 %) and sx unnformed subjects (55 %). The fundamental value seres s the same as for the frst group. Fgure (2) shows the actual tmes seres of the forecasts of the nformed and unnformed partcpants, the equlbrum prces of Group II and the fundamental value of the asset X. 17
Fgure (2). Forecasts of nformed subjects (45 %), unnformed subjects (55 %), equlbrum prces and fundamental value prces n Group II. As we look at these two fgures (1 and 2), the equlbrum prce of the asset X drfts away from the fundamental value n both groups. In addton, the equlbrum prce seres are dfferent. The equlbrum for Group II (equlbrum II) stuck around the ntal value (40 ) whereas the equlbrum for Group I (equlbrum I) showed more of a tendency to follow the fundamental value of the asset X. Accordng to the Augmented Dckey-Fuller tests all seres are non-statonary. To analyze these prce seres further, we transform them to logarthmc returns seres. Table (1) shows the descrptve statstcs of the equlbrum I and equlbrum II returns. The correlaton between the returns seres s 0.21, and the correlaton wth the fundamental value return s 0.68 for equlbrum I and 0.26 for equlbrum II. 18
Table (1). The descrptve statstcs of the seres of returns for equlbrum I and equlbrum II Equlbrum I returns Equlbrum II returns observatons 24 24 mean -0.002 0.002 std. devaton 0.036 0.024 mn -0.080-0.094 max 0.059 0.053 correlaton between 0.206 0.206 Eq I and Eq II correlaton wth 0.683 0.261 fundamental value returns 4.1.3 PANEL TIME SERIES REGRESSION ANALYSIS We have two groups (contanng nne and eleven subjects) and two pools n each group. Thus, we have four dfferent panel data wth T=24. Recall that n the experment the subjects recognze the returns of the varables. To analyze how the subjects n the pools construct ther forecasts we utlze statonary panel tme seres modelng. To be able to control cross-secton correlaton n the ndvdual forecasts, we use the unobserved common factor model. That s, we use common correlated 19
effect modelng [Pesaran 2006]. Because N < T n each estmaton and we have a small sample sze we use the Common Correlated Effect Pooled (CCEP) estmaton method suggested by Pesaran. The CCEP can be computed by OLS regresson: y a bxt d t yt d xt e (2) 1 2 t where the cross-secton means y t T 1 and T k x t are proxes for unobserved common correlated factors. For more nformaton about the methodology of the CCEP, please see Pesaran [2006]. In the set up, we have each ndvdual subject s forecasted return (denoted FO) n the pool at tme t as the dependent varable. The sentment-based beauty contest effect (D), the dfference between the last forecasted return of an ndvdual subject and the last equlbrum return n the group at tme t-1 s our explanatory varable of man nterest. These varables are specfc to each ndvdual. We now have FO t D, (3) t 1 t where t Ct t, 2 and ~.. d. 0, t. The component C t s an ndvdualnvarant tme-specfc unobserved effect. Takng the cross-sectonal averages (A) of equaton (3) gves AFO t ADt 1 C t, (4) t 20
whch we solve for C t as C t 1 AFO t ADt 1 t. (5) The proxes for the common correlated factors (equaton (5)) are the average of the ndvdual subject s forecasted return (AFO) at tme t and the average of the ndvdual subject s beauty contest effect (AD) at tme t-1. These two varables are drectly unobservable for the subjects. In addton, the subjects observe the last equlbrum return (E) n the group at tme t-1 and, for the nformed pools, the fundamental value return of the asset F at tme t. Thus, by nsertng equaton (5), the last equlbrum return n the group (E t-1 ) and the fundamental value return (F t ) nto equaton (3), the nformed (P) pool equaton s PFO t 1 2PDt 1 3Et 1 4Ft ( APFOt APD t 1 t t ), whch gves PFO t * 1 1 t 2PDt 1 3Et 1 4Ft 5 APFO t 6 APD t, (6) where * 1 1, 5, 6, and t t t. For the unnformed (U) pool UFO t * 1 1 t 2UDt 1 3Et 1 4 AUFO t 5 AUDt, (7) 21
where * 1 1, 4, 5, and t t t. Thus, we construct four CCEP regressons (equaton (2)) usng OLS wth followng varables: Dependent varables: Informed I forecasted return t = PFOI t Unnformed I forecasted return t = UFOI t Informed II forecasted return t = PFOII t Unnformed II forecasted return t = UFOII t Explanatory varables: PFOI t-1 EI t-1 = PDI t-1 UFOI t-1 EI t-1 = UDI t-1 PFOII t-1 EII t-1 = PDII t-1 UFOII t-1 EII t-1 = UDII t-1 Equlbrum I return t-1 = EI t-1 Equlbrum II return t-1 = EII t-1 Fundamental value return t = F t Proxes for common correlated factors: Average of the Informed I forecasted return t = APFOI t Average of the Unnformed I forecasted return t = AUFOI t Average of the Informed II forecasted return t = APFOII t Average of the Unnformed II forecasted return t = AUFOII t Average of the PDI t-1 = APDI t-1 22
Average of the UDI t-1 =AUDI t-1 Average of the DPII t-1 = APDII t-1 Average of the DUII t-1 = AUDII t-1 Thus, we estmate the followng OLS regressons: PFOI t * 1 2PDI t 1 t 1 3EI t 1 4Ft 5 APFOI t 6 APDI t, (8) UFOI t * 1 1 t 2UDIt 1 3EI t 1 4 AUFOI t 5 AUDI t, (9) PFOII t * 1 2PDII t 1 t 1 3EII t 1 4Ft 5 APFOII t 6 APDII t, (10) UFOII t * 1 1 t 2UDII t 1 3EII t 1 4 AUFOII t 5 AUDII t, (11) The EMH predcts that n equatons (8) and (10) only β 4 0 wth statstcal sgnfcance and the sentment-based beauty contest phenomenon predcts that n equatons (8), (9), (10) and (11) β 2 0 wth statstcal sgnfcance. In addton, to control the unobservable common correlated varables we estmate the above equatons (8), (9), (10) and (11) wthout common correlated effects. Thus, we estmate the equatons by the LSDV. In all results, we utlze robust standard errors. Table (2). The forecasts of the nformed subjects n the frst group are explaned CCEP coeffcent CCEP p-value LSDV coeffcent LSDV p-value PDI t-1-0.451 0.0000-0.438 0.0000 EI t-1-0.000 1.0000-0.390 0.0000 F t 0.000 1.0000 0.758 0.0000 23
Table (2) shows the coeffcents and p-values of equaton (8), where the ndvdual nformed subject s forecasts n the frst group are regressed. In addton, we estmate the data wth the LSDV. We can learn from the results that the nformed nvestor utlzes her prevous forecast return error (PDI t-1 ) statstcally sgnfcantly n her next forecast. Ths s the sentment-based beauty contest phenomenon. The LSDV estmaton suggests that the common correlated effects (APFOI t and APDI t-1 n equaton (8)) capture the fundamental value returns and prevous equlbrum returns effects n the forecasts. The sentment-based beauty contest effect does not change. Thus we can nterpret ths to mean that the nformed market professonal, n addton to the sentment-based beauty contest effect, utlzes her prvate nformaton (F t ) and last equlbrum return (EI t-1 ) n her forecasts. The results that β 2 0 (and β 3 0) wth statstcal sgnfcance are aganst the EMH. Table (3). The forecasts of the unnformed subjects n the frst group are explaned CCEP coeffcent CCEP p-value LSDV coeffcent LSDV p-value UDI t-1-0.413 0.0000-0.382 0.0000 EI t-1-0.001 0.9960 0.356 0.0013 Table (3) shows the coeffcents and p-values of equaton (9), where the forecasts of the ndvdual unnformed subjects n the frst group are regressed. We can nterpret the results to say that the unnformed nvestor utlzes her prevous forecast return error (UDI t-1 ) statstcally sgnfcantly n her next forecast. Ths s the sentmentbased beauty contest phenomenon. The LSDV estmaton suggests that the common correlated effects (AUFOI t and AUDI t-1 n equaton (9)) capture the prevous equlbrum returns effects n the forecasts. The sentment-based beauty contest effect 24
does not change. Thus we can nterpret ths to mean that the unnformed market professonal, n addton to the sentment-based beauty contest effect, utlzes her last equlbrum return (EI t-1 ) n her forecasts. The result that β 2 0 (and β 3 0) wth statstcal sgnfcance s aganst the EMH. Table (4). The forecasts of the nformed subjects n the second group are explaned CCEP coeffcent CCEP p-value LSDV coeffcent LSDV p-value PDII t-1-0.473 0.0000-0.471 0.0000 EII t-1-0.000 1.0000-0.547 0.2576 F t 0.000 1.0000 0.229 0.1290 Table (4) shows the coeffcents and p-values of equaton (10), where the forecasts of the ndvdual nformed subjects n the second group are regressed. We can nterpret the results to say that the nformed nvestor utlzes her prevous forecast return error (PDII t-1 ) statstcally sgnfcantly n her next forecast. Ths s the sentment-based beauty contest phenomenon. The LSDV estmaton suggests exactly the same results. The sentment-based beauty contest effect does not change. Thus, the results suggest that when the nformed market professonal s n the mnorty n the market she does not utlze her prvate nformaton about the true fundamental value. The result that β 2 0 wth statstcal sgnfcance s aganst the EMH. Table (5). The forecasts of the unnformed subjects n the second group are explaned CCEP coeffcent CCEP p-value LSDV coeffcent LSDV p-value UDII t-1-0.143 0.1037-0.205 0.0019 EII t-1 0.003 0.9952 0.690 0.1122 Table (5) shows the coeffcents and p-values of equaton (11), where the forecasts of the ndvdual unnformed subjects n the second group are regressed. The CCEP 25
results show that the unnformed nvestor behaves as a martngale. The LSDV estmaton suggests that the common correlated effects (AUFOII t and AUDII t-1 n equaton (11)) capture the sentment-based beauty contest effects n the forecasts. We can nterpret ths to say wth cauton that the unnformed market professonal, when she s n the majorty n the market, demonstrates the sentment-based beauty contest effect. The CCEP suggests that when the unnformed market professonal s n the majorty n the market she behaves as the EMH predcts, but the LSDV suggests that the sentment-based beauty contest effect s present. 5. CONCLUSIONS We tested the behavor of stock market professonals n an expermental settng. We utlzed artfcal data from the fnancal markets (random walk wthout a drft) as the fundamental value of the asset, and asked the stock market professonals to forecast the equlbrum prce at the next step ahead wth the avalable nformaton. We splt twenty experenced market professonals nto two groups, n whch the rato of nformed subjects to unnformed subjects was dfferent, and we splt the groups nto two pools (nformed and unnformed subjects). One subject pool n the group receved the prce change of the fundamental value at the next step ahead (prvate nformaton) n addton to the realzed equlbrum prces (publc nformaton). Another pool n the group were purely nose traders whose only nformaton was the realzed equlbrum prces. Thus we have a short horzon (one perod) forecastng problem to solve. Accordng to the EMH, even n a very short horzon nvestng perod a ratonal nvestor should behave as a long horzon agent wth an nfnte lfe. That s, she should speculate on long horzon fundamentals. In our settng, the 26
fundamental value of the asset followed a random walk wthout drft, and the nformed subjects knew that value at the next step ahead. The EMH predcts that an nformed nvestor wth knowledge of the true value of the asset should utlze that nformaton and entrely gnore past prce changes. Keynes noted [1936] that nvestors care about hgher order belefs. That s, they try to estmate what other nvestors are dong. Keynes calls ths phenomenon a beauty contest n fnancal markets. In our expermental settng, the sentment-based beauty contest effect s the observable (by the subject) dfference between her last forecasted return and the last equlbrum return, because the return dfference s her forecasted error from the prevous prce change. Ths s sentment-based because the past returns should not affect the next returns accordng to the EMH. In the frst group 56 % were nformed subjects (wth prvate and publc nformaton) and 44 % were unnformed subjects, and n the second group 45 % were nformed and 55 % were unnformed subjects. In the analyss of the results, we found, wth statstcal sgnfcance, that n all the groups the market professonals demonstrated the beauty contest phenomenon. The results ndcate that the sentment-based beauty contest effect occurs wth a 0,001 % sgnfcance level n three out of the four pools. Only when the unnformed professonals are n the majorty n market are the results mxed. The CCEP suggests that an unnformed professonal behaves as the EMH predcts, but the LSDV suggests that the sentment-based beauty contest effect s present. In addton, the results suggest that when the nformed professonal s n a majorty n the market, she also utlzes her prvate nformaton about the true value 27
of the asset. The results are consstent wth the theoretcal beauty contest lterature. In both cases, the equlbrum prces drft away from the fundamental value. REFERENCES Abreu, D., and M. Brunnermeer [2003]. Bubbles and crashes, Econometrca, 17, 173 204. Alevy, J., Hagh, M., and J. Lst [2007]. Informaton cascades: Evdence from a feld experment wth fnancal market professonals. Journal of Fnance, 112, 151 180. Allen, F., Morrs S., and H. Shn [2006]. Beauty contest and terated expectatons n asset markets. Revew of Fnancal Studes, 19, 719 752. Anderson, L., and C. Holt [1997]. Informaton cascades n the laboratory. Amercan Economc Revew, 87, 847 862. Bacchetta, P., and E. van Wncoop [2008]. Hgher order expectatons n asset prcng. Journal of Money, Credt and Bankng, 40, 837 866. Banerjee, S., Kanel, R., and I. Kremer [2009]. Prce drft as an outcome of dfferences n hgher-order belefs. Revew of Fnancal Studes, 22, 3707 3734. Cpran, M., and A. Guarno [2005]. Herd behavor n a laboratory fnancal market. Amercan Economc Revew, 95, 1427 1433. 28
Cpran, M., and A. Guarno [2008]. Herd behavor and contagon n fnancal markets. B. E. Journal of Theoretcal Economcs, 8, 1 56. Cpran, M., and A. Guarno [2009]. Herd behavor n fnancal markets: An experment wth fnancal market professonals. Journal of European Economc Assocaton, 7, 206 233. Drehmann, M., Oechssler J., and A. Roder [2005]. Herdng and contraran behavor n fnancal market: An nternet experment. Amercan Economc Revew, 95, 1403 1426. Froot, K., Scharfsten D., and J. Sten [1993]. Herd on the street: nformatonal neffcences n a market wth short-term speculaton, Journal of Fnance, 77, 1461 1484. Hagh, M., and J. Lst [2005]. Do professonal traders exhbt myopc loss averson? An expermental analyss. Journal of Fnance, 110, 523 534. Hagh, M., and J. Lst [2010]. Investment under uncertanty: testng the optons model wth professonal traders. Revew of Economcs and Statstcs, 92, 974 984. Hommes, C., Sonnemans, J., Tunstra, J., and H. van de Velden [2005]. Coordnaton of expectatons n asset prcng experments. Revew of Fnancal Studes, 18, 955 980. 29
Ilomäk, J. [2012]. Framed feld experment wth stock market professonals. Journal of Behavoral Fnance, 13, 251 258. Keynes, J. [1936]. General theory of nterest, employment and money. London McMllan. Kocher, M., and M. Sutter [2005]. The Decson maker matters: ndvdual versus group behavour n expermental beauty-contest games. Economc Journal, 115, 200 225. Nagel, R. [1995]. Unravelng n guessng games: an expermental study. Amercan Economc Revew, 85, 1313 1326. Pesaran, H. [2006]. Estmaton and nference n large heterogeneous panels wth a multfactor error structure. Econometrca, 74, 96 1012. 30
APPENDIX I: METHODOLOGY OF THE EXPERIMENT 1. The subjects are experenced stock market professonals. 2. There are two groups and two pools n both groups (56 % nformed and 44 % unnformed) and (45 % nformed and 55 % unnformed) and the subjects are chosen randomly. 3. The subjects are seated such that they make ther decsons ndependently. The communcaton between the subjects s prohbted. 4. The subjects can use ther computers to calculate returns and execute analyses. 5. The task of the subject s the followng: to forecast next step ahead prce of the unspecfed asset X n the whole number accuracy wth the avalable nformaton. 6. There are 24 forecast to be made for every subject. 7. The subject has her 24 answers sheets and she wrte down ndependently her forecast after she has told to do so n every step. 8. The subject that has selected to be n the nformed pool gets her prvate nformaton n the answerng sheet. 9. The prvate nformaton s relable nformaton and that s told to the subjects. 10. The common nformaton (equlbrum prce of the asset X) s revealed after the subjects have made ther forecast. Ths leads to the fact that every subject recognzes her own forecastng accuracy as the experment goes on. 11. The subjects can wrte down the hstorcal nformaton as the experment goes on. 12. The subject earns that realzed return (1% = 1 ) f the sgn of the return s correct and she lose that return f the sgn s ncorrect. If the sgn and the whole number are correct, the subject earns the return on double. 31