The disposition effect and investor experience

Transcription

1 MPRA Munch Personal RePEc Archve The dsposton effect and nvestor experence ewton Da Costa Jr and Marco Goulart and Cesar Cupertno and Jurandr Macedo Jr and Sergo Da Slva Federal Unversty of Santa Catarna 2013 Onlne at MPRA Paper o , posted 4. January :05 UTC

2 The dsposton effect and nvestor experence ewton Da Costa Jr. a,b*, Marco Goulart b, Cesar Cupertno b, Jurandr Macedo Jr. c, Sergo Da Slva a a Graduate Program n Economcs, Federal Unversty of Santa Catarna, , Floranopols SC, Brazl b Graduate Program n Busness, Federal Unversty of Santa Catarna, , Floranopols SC Brazl c Department of Knowledge Engneerng, Federal Unversty of Santa Catarna, , Floranopols SC, Brazl Abstract We examne whether nvestng experence can dampen the dsposton effect, that s, the fact that nvestors seem to hold on to ther losng stocks to a greater extent than they hold on to ther wnnng stocks. To do so, we devse a computer program that smulates the stock market. We use the program n an experment wth two groups of subjects, namely experenced nvestors and undergraduate students (the nexperenced nvestors). As a control procedure, we consder random trade decsons made by robot subjects. We fnd that though both human subjects show the dsposton effect, the more experenced nvestors are less affected. JEL classfcaton: G11; G14 Keywords: Dsposton effect; Investor experence; Artfcal stock market; Framed feld experment 1. Introducton The dsposton effect s the anomaly that nvestors seem to hold on to ther losng stocks to a greater extent than they hold on to ther wnnng stocks (Schlarbaum et al., 1978, Shefrn and Statman, 1985, Weber and Camerer, 1998). For nstance, data from a consultng retal brokerage house revealed that stocks wth postve returns were 68 percent more lkely to be sold than those wth negatve returns (Odean, 1998). The dsposton effect s lessened f there s fnancal counselng (Taylor, 2000, Shapra and Veneza, 2001), and t s heghtened for nexperenced nvestors (Grnblatt and Keloharju, 2001, Coval and Shumway, 2005, Feng and Seasholes, 2005, Locke and Mann, 2005, Dhar and Zhu, 2006), though that s stll unsettled (Chen et al., 2007). Here, we nvestgate the relatonshp between the dsposton effect and nvestng experence usng a framed feld experment (Harrson and Lst, 2004). Tests provng the dsposton effect n actual markets (such as those n the works above) cannot be conclusve because nvestor decsons cannot be controlled n there. For that reason, lab experments can be more llumnatng n that they can be desgned to match ndvdual nvestors tradng decsons wth the prces at whch they buy or sell stocks. In stark contrast wth the above studes usng actual data, when t comes to the lab the dsposton effect may be even hgher for experenced nvestors (lke n the artefactual feld experments of Hagh and Lst, 2005, and of Abbnk and Rockenbach, 2006). That can be explaned by ether the curse of knowledge ( the more you know, the worse you become at

3 usng that knowledge ) (Camerer et al., 1989), the desre to avod regret (Barber and Odean, 1999), or smply by the fact that an experment s too smplstc. Because t s possble that the relatonshp between the dsposton effect and nvestng experence can be dependent on experment desgn, here we try to remedy such a defcency by developng a computer program that mmcs the stock market whle retanng the characterstc that nvestor decsons cannot nfluence the (exogenous) stock prces. We use the program n an experment wth two groups of subjects, namely experenced nvestors and undergraduate students (the nexperenced nvestors). As a control procedure, we also consder random trade decsons made by robot subjects. We thus set a more complex expermental envronment than does a typcal experment whle preservng the control characterstcs that are the edge of the expermental method. As a result, we fnd the dsposton effect n human subjects, and also that experenced nvestors are less prone to the effect, whch s n lne wth most of the evdence dscussed above for actual data. Harrson and Lst (2004) put forward the followng taxonomy to classfy experments: (1) conventonal lab experment; (2) artefactual feld experment; (3) framed feld experment; and (4) natural feld experment. As observed, ours s a framed feld experment, whch s also an artefactual feld experment but wth feld context n the task and nformaton set used by the subjects. Table 1 presents the man results of selected recent work related to the dsposton effect and nvestor experence; the reader may wsh to consder the references theren for a more comprehensve account of the vast lterature on the subject. The rest of ths paper s organzed as follows. Secton 2 presents the three measures of the dsposton effect employed n ths work, Secton 3 detals the desgn of the experment, Secton 4 presents the characterstcs of the subjects partcpatng n the experment, Secton 5 reports results, and Secton 6 concludes the study. A senstvty analyss of the results s presented n an Appendx. 2. Measures of the dsposton effect Expermental studes typcally track the dsposton effect whenever subjects sell more (less) stocks as the sale prce s above (below) ether the purchasng prce or the prevous prce (Weber and Camerer, 1998). However, such a measure can be msgudng n the presence of bull-bear market cycles. For nstance, n a bull market a stock sold s more lkely to be a wnner. Here nvestors mght ratonally thnk that rsng prces wll tend to persst n future, thereby makng sense to sell wnners (Da Costa Jr et al., 2008). Snce our experment s run n an artfcal market we consder the measure of the dsposton effect commonly used n realworld markets (Odean, 1998) that s able to take market cycles nto account. However, Odean s measure s not wthout problems, as dscussed below. For that reason, we also assess the dsposton effect n our experment by two other measures: that of Weber and Camerer (1998), and a more recent one suggested by Dhar and Zhu (2006). Odean s measure consders the actual and potental trades of nvestor durng a sample perod. Potental trades refer to stocks n a portfolo that were not sold but that could have been ether wnners or losers. The proporton of gans realzed ( PGR ) and proporton of losses realzed ( PLR ) are computed as lr PGR =, PLR = (1) + + gr gr gp lr lp

4 where gr ( lr ) s the number of trades by nvestor wth a realzed gan (loss), and ( lp ) s the number of potental trades for nvestor wth a gan (loss). The dsposton effect ( DE ) of nvestor s then gp DE = PGR PLR (2) where 1 DE 1. A postve value of DE ndcates that a smaller proporton of losers s sold f compared wth the proporton of wnners sold, n whch case nvestor exhbts the dsposton effect. The defnton n equaton (2) can be evaluated by the t-statstc t PGR PLR = (3) SE where the standard error SE s PGR(1 PGR) PLR(1 PLR) SE = gr gp lr lp. (4) One dsadvantage of equaton (2) s that the PGR and PLR measures are senstve to portfolo sze and tradng frequency (Odean, 1998). They are lkely to be smaller for nvestors who hold larger portfolos and trade frequently because those portfolos contan a larger number of stocks wth captal gans and captal losses. Ths problem gets more serous as the measures are employed n cross-sectonal analyses. Thus we also employ two other measures of the dsposton effect that are not senstve to portfolo sze and tradng frequency. The frst one s precsely the measure of Weber and Camerer (1998), whch consders the dfference between the number of trades wth realzed gans by nvestor and the number of trades wth realzed losses relatve to the number of all trades, that s, DE gr lr = gr + lr (5) where 1 DE 1. If the number of trades wth realzed gans matches the number of trades wth realzed losses there s no dsposton effect. The other measure s that of Dhar and Zhu (2006): DE =. (6) gr gp lr lp 3. Experment desgn To run our experment we employ the computer program that smulates the stock market called SmulaBolsa, whch was developed by one of us (J.M.). Fgure 1 shows the program s man menu. The program generates an ndvdual report for all the decsons made

5 by the subjects throughout the smulaton perod. The output can thus allow one to get nformed about varables, such as the number of stocks bought and sold each perod, and ndvdual portfolo composton at the end of a perod. The program was fed wth actual data for stock prces taken from the Sao Paulo stock exchange (Bovespa) for the fve-year perod from January 1997 to December The program also ncluded ndcators based on fundamental analyss taken from Economatca. However, subjects were not nformed about the perod nvolved, and the companes real names were replaced wth fcttous ones. The stock prces were deflated by the Brazlan GDP deflator (called IGP-DI), and corrected for dvdends and other occurrences. Then the prces were normalzed so that each stock cost one Brazlan real (R$ 1) at the begnnng of the experment. Because prces (purchasng prces equalng sellng prces) were fxed by the smulator, the whole stock market was exogenous to each subject. Each subject was then consdered as a small trader, and ther actons dd not nfluence prces. On the man screen of the program (Fgure 1) each subject was endowed wth ntal assets worth R$ 300,000, whch could be allocated n ether cash or stocks of 28 companes. Money for real was not nvolved. Before gettng started, subjects receved nstructons about the smulaton. They were told a story about managng the portfolo of a teenage boy s best frend who ded of cancer. The story amed at promotng emotonal engagement wth the play, to compensate for the fact that no real money was nvolved. Subjects were then asked to manage the nvestment portfolo over 20 perods usng the program, and ther buy and sell decsons had to be made at the begnnng of each perod. Decsons were to be reached wthn a three mnute tme lmt. After ths lmt, the smulaton screen swtched for the next one. The subjects could eventually compare ther decsons durng the experment wth the actual stock prces announced by the program. Subjects performed a total 7,429 transactons, whch s equvalent to fve years of actual data (Table 2). 4. Subject characterstcs Stock nvestors wth a mnmum of two years of experence were sampled from consultng retal brokerage houses located n Floranopols, Brazl. From the 26 subjects that ended up partcpatng n the experment, 9 reported more than fve years of experence, whle 17 reported from two to fve years. The nexperenced nvestors were sampled from economcs and busness admnstraton students of the Federal Unversty of Santa Catarna, also located n Floranopols. The students had already taken Captal Markets n the prevous term. A total 38 student subjects partcpated. The experment wth students was conducted durng two sessons run n the second term of The sessons were located at the unversty s Stock Market Lab. The Lab has 40 computers arranged n ndvdual tables wth no possble communcaton between users. Thus, one subject s screen could not be seen by others. The students sessons took approxmately 90 mnutes each. The experment wth professonal nvestors was run n the course of several sessons performed n ther own workplaces to comply wth ther tme avalablty. The sessons took 90 mnutes as well. In all experments, subjects were allowed to ask for drectons from the tutor. After the end of the sessons we calculated the measures of the dsposton effect as n Secton 2. To control the experment, we also consdered 50 robot subjects that were programmed to randomly buy or sell stocks through a unform dstrbuton (see Mller, 2008 and references theren). Then we calculated the dsposton effect for the robots as well. Ths procedure amed at checkng whether the effect was really caused by some type of cogntve lluson of the human bran, as commonly asserted. After all, f robots also exhbted the dsposton effect t should be explaned by some type of emergence property resultng solely

6 from the dynamcs of the experment; n other words, the effect had nothng to do wth cogntve lluson. The fnal sample of 26 nvestors and 38 students dd not nclude those subjects wth nether gans nor losses throughout the experment, and also those that spent less than 30 mnutes n the smulaton. We thought that such subjects were not really engaged wth the experment. 5. Results Table 2 shows the descrptve statstcs of the three types of subjects. As can be seen, the average number of trades carred out by the nvestors was less than that of students (dfference = 8.8, t = 1.53, p-value = 0.13). For both human subjects the total average returns beat the actual returns n the Bovespa ndex (over the perod January 1997 to December 2001). Moreover, the returns of human subjects were by far greater than the returns made by the robots. Table 3 shows the detals of the calculaton of the dsposton effect usng equaton (2) for each of the three groups of subjects as a whole. The t -statstc for both groups of human subjects was t = Ths fgure matches those commonly found n the studes wth actual data descrbed n the Introducton. As can be seen n Table 3, though both human groups exhbted the dsposton effect, the effect was lessened for the experenced nvestors. The robots dd not show the dsposton effect, thus suggestng that the common explanaton by some type of human cogntve bas makes sense. Table 4 shows the descrptve statstcs for the dsposton effect calculated separately for each ndvdual usng equaton (2). It also shows two tests (parametrc and nonparametrc) for the dstrbuton of the effect along wth a test for the normalty of the dstrbuton (Jarque- Bera). The t-statstc tests the hypothess of zero mean, whle the nonparametrc Wlcoxon Z- statstc tests whether the medan of the dstrbuton s zero. Table 5 repeats the calculatons consderng the defnton n equaton (5), whle Table 6 consders the measure represented by equaton (6). As can be seen, the results n Tables 4 6 are very smlar. In Tables 4 and 6, the dsposton effect s sgnfcant at one percent for students and students and nvestors taken together, and sgnfcant at fve percent for the nvestors alone. Yet the effect s overall sgnfcant at one percent n Table 5. Apart from the nvestors n the defnton (6) n Table 6, the Jarque-Bera test could not reject the normalty hypothess. For the robots, the hypothess could be rejected (Tables 4 and 6), but such subjects dd not exhbt the dsposton effect, as seen. The row at the bottom n Table 6 also shows that a lttle bt more than 26 percent of the subjects dd not exhbt the dsposton effect; ths fndng matches that of Dhar and Zhu (2006) for actual data, where 20 percent of ndvduals dd not present the effect. To nvestgate the nfluence of experence on the dsposton effect, we frst run a smple lnear regresson between the dsposton effect (usng the defntons n equatons (2), (5), and (6)) and nvestng experence, that s, DE = α + β X + µ (7) where experence s tracked by the dummy varable X, whch takes on the value X = 1 for the subjects wth two or more years n stock markets (nvestors), and X = 0 for subjects wth experence below two years (students). Table 7 shows that the dsposton effect s reduced as the years of experence grow (negatve slope coeffcent); however, the coeffcent s nonsgnfcant.

7 To remedy such a shortcomng, we run the followng multple regresson: DE = α + β X + β X + µ (8) where X 1 = 1 s the dummy for subjects wth 2 to 5 years of experence ( X 1 = 0 otherwse), and X 2 = 1 s the dummy for subjects wth more than 5 years of experence ( X 2 = 0 otherwse). Table 8 shows that the dsposton effect tends to be reduced for subjects wth more than fve years n stock markets. Parameter β 2 was sgnfcant at fve percent for the defntons n equatons (5) and (6), and was sgnfcant at 10 percent for the measure gven by equaton (2). Moreover, we test for the dfference between the dsposton effects of the two experenced groups only. Usng the defntons n equaton (2), (5), and (6), we then run the followng regresson: DE = α + β X + µ, (9) where the level of experence s tracked by the dummy varable X, whch takes on the value X = 1 for the subjects wth more than 5 years n stock markets, and X = 0 for the subjects wth 2 to 5 years of experence. Table 9 shows that the dfference between the two groups was sgnfcant. The slope coeffcent was negatve for all the three dfferent defntons of the dsposton effect. It was sgnfcant at fve percent for the defnton n equaton (2), 10 percent for that n equaton (5), and one percent for that n equaton (6). 6. Concluson Does nvestor experence dampen the dsposton effect? Most studes usng actual data answer yes. However, ths answer conflcts wth the results found n lab experments. Tests provng the dsposton effect n actual markets cannot be conclusve because nvestor decsons cannot be controlled n there. Control characterstcs are an advantage of the expermental method, though results can stll be dependent on experment desgn, manly f the experment s too smplstc. Here, we consder that possblty and thus devse a more elaborated experment through a computer program that mmcs the stock market whle retanng the control characterstcs. In lne wth the actual data studes, our framed feld experment found that the dsposton effect s reduced f nvestors have more than fve years of experence n stock markets. The dsposton effect s commonly attrbuted to a cogntve lluson of the human bran. To evaluate such a proposton we consder not only professonal nvestors and students n our experment, but also robots. Do robots dream of wnnng stocks? If they do, the phenomenon has to be explaned by some type of emergence property resultng from the dynamcs of the experment, rather than by human bran mperfectons. We fnd that robots do not exhbt the effect, and thus we cannot dsmss that the phenomenon s really caused by cogntve lluson.

8 Fgure 1. Man menu of the stock market smulator we devsed to run the experment

9 Table 1. Selected recent work related to the dsposton effect and nvestor experence Author Result Menkhoff and kforow (2009) Fund managers who have strong ncentves to learn effcent behavor and who do not endorse the behavoral fnance vew, end up falng to learn, thus suggestng that many behavoral fnance patterns are rooted n human behavor and dffcult to be overcome by learnng Chang (2008) Klger and Kudryavtsev (2008) Lee et al. (2008) Goetzmann and Massa (2008) Hales (2007) Hedesstrom et al. (2007) Garvey and Murphy (2004) Evdence of the dsposton effect n nvestors of the Tawanese warrant markets The reference pont updatng process of the dsposton effect s more reactve to events when nformaton flow s low and prces are senstve to market fluctuatons. Agents facng numerous alternatves consder those that have caught ther attenton Evdence of the dsposton effect n nternet-based stock tradng A panel of ndvdual nvestor tradng records shows that exposure to a portfolo of stocks held by dsposton-prone nvestors explans cross-sectonal dfferences n daly returns Investors are motvated to agree unthnkngly wth nformaton that suggests they mght make money on ther nvestment, but dsagree wth nformaton that suggests they mght lose money In an nternet-based survey of fcttous choces among fund categores, home bas and a dversfcaton heurstc were unaffected by prevous stock market nvestment experence Data on a U.S. propretary stock-tradng team provde evdence of the dsposton effect Table 2. Descrptve statstcs of the subjects partcpatng n the smulaton Subject Investors Students Robots All subjects Sample Average stocks n portfolo Total trades 1,644 2,737 3,048 7,429 Average number of trades Cumulatve returns, % Bovespa ndex cumulatve returns, % 178.3

10 Table 3. The dsposton effect for the groups of subjects usng equaton (2) Subject Group of nvestors Group of students Both human groups Group of robots gr lr gp lp PGR PLR DE SE t -statstc 5.51 *** *** *** 0.52 *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Table 4. The dsposton effect for ndvdual subjects usng equaton (2) Subject Investors Students All human subjects Robots umber of subjects Mean Medan Maxmum Mnmum Standard devaton Jarque-Bera ** t-statstc (mean = 0) 2.57 ** 3.26 *** 4.19 *** 1.15 Wlcoxon Z-statstc (medan = 0) 2.18 ** 3.21 *** 3.93 *** 0.74 Subjects wth DE > 0, % *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Table 5. The dsposton effect for ndvdual subjects usng equaton (5) Subject Investors Students All human subjects Robots umber of subjects Mean Medan Maxmum Mnmum Standard devaton Jarque-Bera t-statstc (mean = 0) 4.25 *** 7.91 *** 8.60 *** 0.41 Wlcoxon Z-statstc (medan = 0) 3.34 *** 4.90 *** 5.93 *** 0.28 Subjects wth DE > 0, % *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Table 6. The dsposton effect for ndvdual subjects usng equaton (6) Subject Investors Students All human subjects Robots umber of subjects Mean Medan Maxmum Mnmum Standard devaton Jarque-Bera *** *** *** t-statstc (mean = 0) 2.45 ** 2.59 ** 3.45 *** 0.48 Wlcoxon Z-statstc (medan = 0) 2.08 ** 3.21 *** 3.76 *** 0.12 Subjects wth DE > 0, % *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1%

11 Table 7. The dsposton effect and nvestng experence (equaton (7)) Dsposton effect Equaton (2) Equaton (5) Equaton (6) Constant *** *** *** Experence ( 2 years) R squared n = *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Table 8. The dsposton effect and nvestng experence (equaton (8)) Dsposton effect Equaton (2) Equaton (5) Equaton (6) Constant *** *** ** Experence (2 5 years) Experence (> 5 years) * ** * R squared n = *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Table 9. The dsposton effect between the two groups of experenced subjects (equaton (9)) Dsposton effect Equaton (2) Equaton (5) Equaton (6) Constant *** *** *** Experence (> 5 years) ** * ** R squared n = *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1%

12 References Abbnk, K., Rockenbach, B., Opton prcng by students and professonal traders: A behavoural nvestgaton. Manageral and Decson Economcs 27, Barber, B.M., Odean, T., The courage of msguded convctons. Fnancal Analysts Journal 55, Camerer, C., Loewensten, G., Weber, M., The curse of knowledge n economc settngs: An expermental analyss. Journal of Poltcal Economy 97, Chang, C.-H., The mpact of behavoral ptfalls on nvestors decsons: The dsposton effect n the Tawanese warrant market. Socal Behavor and Personalty 36, Chen, G.M, Km, K., ofsnger, J., Ru, O., Tradng performance, dsposton effect, overconfdence, representatveness bas, and experence of emergng market nvestors. Journal of Behavoral Decson Makng 20, Coval, J.D., Shumway, T., Do behavoral bases affect prces? Journal of Fnance 60, Dhar, R., Zhu,., Up close and personal: Investor sophstcaton and the dsposton effect. Management Scence 52, Da Costa Jr,., Mneto, C., Da Slva, S., Dsposton effect and gender. Appled Economcs Letters 15, Feng, L., Seasholes, M.S., Do nvestor sophstcaton and tradng experence elmnate behavoral bases n fnancal markets? Revew of Fnance 9, Garvey, R., Murphy, A., Are professonal traders too slow to realze ther losses? Fnancal Analysts Journal 60, Goetzmann, W.., Massa, M., Dsposton matters: Volume, volatlty, and prce mpact of a behavoral bas. Journal of Portfolo Management 34, Grnblatt, M., Keloharju, M., What makes nvestor trade? Journal of Fnance 56, Hagh, M.S., Lst, J.A., Do professonal traders exhbt myopc loss averson? An expermental analyss. Journal of Fnance 60, Hales, J., Drectonal preferences, nformaton processng, and nvestors forecasts of earnngs. Journal of Accountng Research 45, Harrson, G.W., Lst, J.A., Feld experments. Journal of Economc Lterature 42,

13 Hedesstrom, T.M., Svedsater, H., Garlng, T., Determnants of the use of heurstc choce rules n the Swedsh Premum Penson Scheme: An Internet-based survey. Journal of Economc Psychology 28, Klger, D., Kudryavtsev, A., Reference pont formaton by market nvestors. Journal of Bankng & Fnance 32, Kumar, A., Lm, S.S., How do decson frames nfluence the stock nvestment choces of ndvdual nvestors? Management Scence 54, Lee, H.-J., Park, J., Lee, J.-Y., Wyer Jr, R.S., Dsposton effects and underlyng mechansms n e-tradng of stocks. Journal of Marketng Research 45, Locke, P.R., Mann, S.C., Professonal trader dscplne and trade dsposton. Journal of Fnancal Economcs 76, Menkhoff, L., kforow, M., Professonals endorsement of behavoral fnance: Does t mpact ther percepton of markets and themselves? Journal of Economc Behavor & Organzaton 71, Mller, R.M., Don t let your robots grow up to be traders: Artfcal ntellgence, human ntellgence, and asset-market bubbles. Journal of Economc Behavor & Organzaton 68, Odean, T., Are nvestors reluctant to realze ther losses? Journal of Fnance 53, Schlarbaum, G.G., Lewellen, W.G., Lease, R.C., Realzed returns on common stocks nvestments: The experence of ndvdual nvestor. Journal of Busness 51, Shapra, Z., Veneza, I., Patterns of behavor of professonally managed and ndependent nvestors. Journal of Bankng & Fnance 25, Shefrn, H., Statman, M., The dsposton to sell wnners too early and to rde losers too long: Theory and evdence. Journal of Fnance 40, Taylor, L., The dsposton effect: Do ew Zealand nvestors keep ther mstakes? Coursework Masters Thess n Fnance, Unversty of Otago. Weber, M., Camerer, C.F., The dsposton effect n securtes tradng: An expermental analyss. Journal of Economc Behavor & Organzaton 33,

14 Appendx. Senstvty analyss We conducted a slghtly modfed experment setup two years tme after the frst experment descrbed n the man text. Its major features were: 10 assets were consdered, rather than the 28 assets used n the frst experment; each subject was endowed wth R$ 300,000 at the begnnng of the experment; all stock prces and company names were dfferent from the frst experment s prces and names; as n the frst experment, subjects who made nether gans nor losses throughout the experment were dropped from the sample; those who spent less than 10 mnutes n the experment were excluded from the sample; as n the frst experment, the sample of nexperenced nvestors was drawn from students of busness admnstraton of the Federal Unversty of Santa Catarna. The students had already coursed Captal Markets n the prevous term. The experment wth the students was conducted n one sesson run durng the frst term of 2009; the experment wth professonal nvestors was run n the course of several sessons that were conducted n ther own workplaces to comply wth ther tme avalablty, durng the second semester of 2009; stock nvestors wth a mnmum of two years of experence were sampled from two dfferent consultng retal brokerage houses located n Floranopols, Brazl. From the 20 subjects that ended up partcpatng n the experment, sx reported more than fve years of experence, whle ten reported two- to fve years of experence. Four subjects were excluded from the sample because one dd not make any sell transacton and three belatedly reported to have less than two years of experence. Also, one subject msreported to have more than fve years of experence, but n fact he had only three years. So, he was reallocated from the more than fve years of experence to the twoto fve years of experence group. The results were smlar to those of the benchmark experment, and are descrbed n the tables below. Shorter samples were responsble for less statstcal sgnfcance, however. The dsposton effect was nonsgnfcant among the nvestors, both as a group and ndvdually (Tables 2A and 4A). Despte that, the effect stll affected less the nvestors (negatve slope). Also, Table 5A shows that the more experenced nvestors were less affected by the dsposton effect f compared wth the nvestors wth experence between two- to fve years (negatve slope). In order to drectly compare the two groups of nvestors we run a smple regresson based n equaton (9), as n the benchmark experment. The results n Table 6A show that the dfference between the two groups of nvestors s sgnfcant. The group of nvestors wth less market experence (two- to fve years) presents a dsposton effect of , whle n the more experenced group (> fve years) the coeffcent (whch represents the dfferental effect) s and sgnfcant at 10%. We also run a smple lnear regresson between the dsposton effect (as n equaton (2)) and years of nvestng experence, nstead of usng dummes as n equatons (7) and (8). We show the results n Table 7A. In the benchmark experment we dd not collect such data (subjects were only checked as to whether they had no experence, more than two years, or more than fve years of experence). We detect the presence of heteroskedastcty n the varance-covarance matrces of the regressons presented n Tables 4A, 5A, 6A, and 7A, but the problem was properly corrected by the technque of Whte.

15 Table 1A. Descrptve statstcs of the subjects partcpatng n the extra smulaton Subject Investors Students Robots All human subjects Sample Total trades Average number of trades Table 2A. The dsposton effect for the groups of subjects usng equaton (2) Subject Group of nvestors Group of students Both human groups Group of robots gr lr gp lp PGR PLR DE SE t -statstc *** 3.13 *** 0.56 *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Table 3A. The dsposton effect for ndvdual subjects usng equaton (2) Subject Investors Students All human subjects Robots umber of subjects Mean Medan Maxmum Mnmum Standard devaton Jarque-Bera *** t-statstc (mean = 0) * 2.09 ** 0.72 Wlcoxon Z-statstc (medan = 0) * 1.64 * 0.43 Subjects wth DE > 0, % *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Table 4A. The dsposton effect and nvestng experence (equaton (7)) Dsposton effect Equaton (2) Constant * (p-value = 0.06) Experence ( 2 years) (p-value = 0.19) R squared *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Regresson wth Whte heteroskedastcty-consstent standard errors and covarance, n = 37 Table 5A. The dsposton effect and nvestng experence (equaton (8)) Dsposton effect Equaton (2) Constant * (p-value = 0.06) Experence (2 5 years) (p-value = 0.49) Experence (> 5 years) ** (p-value = 0.02) R squared *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Regresson wth Whte heteroskedastcty-consstent standard errors and covarance, n = 37

16 Table 6A. The dsposton effect between the two groups of experenced subjects (equaton (9)) Dsposton effect Equaton (2) Constant (p-value=0.24) Experence (> 5 years) * (p-value=0.09) R squared *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Regresson wth Whte heteroskedastcty-consstent standard errors and covarance, n = 16 Table 7A. The dsposton effect and years of nvestng experence Dsposton effect Equaton (2) Constant ** (p-value = 0.04) Years of experence * (p-value = 0.07) R squared *sgnfcant at 10%, **sgnfcant at 5%, ***sgnfcant at 1% Regresson wth Whte heteroskedastcty-consstent standard errors and covarance, n = 37