Bubbles, Experience, and Success

Transcription

1 Bubbles, Experience, and Success Dmitry Gladyrev, Owen Powell, and Natalia Shestakova March 15, 2015 Abstract One of the most robust findings in experimental asset market literature is the experience effect asset markets populated by traders who are familiar with the market environment demonstrate a very high level of price efficiency. A few studies suggest that the effect still holds even in mixed-experience markets, i.e. when some of the experienced traders are replaced with inexperienced. Results of our asset market experiment, standard in most aspects, suggest that, first, the experience effect is not as robust as was previously thought and, second, price efficiency of mixed-experience markets is sensitive to the previous success of experienced traders. Specifically, mixed-experience markets with experienced traders being least and most successful in the past are characterized by a lower price efficiency than those with experienced traders being moderately successful. One general law, leading to the advancement of all organic beings, namely, multiply, vary, let the strongest live and the weakest die. Charles Darwin, 1859 On the Origin of Species The market selection hypothesis, dating back at least to Alchian (1950), posits that over time natural selection happens in markets: more successful traders survive, whereas those with trading losses disappear. If success is at least somewhat determined by behavior, then as a result of market selection traders who survive in markets behave differently than traders who are randomly drawn from the population to participate in markets. The effect of market selection on market outcomes is hard to evaluate empirically, since the counterfactual, i.e. the behavior of traders randomly drawn from the population, is not observable. One market outcome that is particularly important is market efficiency, as efficient markets produce prices that are accurate indicators of relative values. The actual efficiency of real markets has long been debated (Fama, 1970; Lim and Brooks, 2011) since underlying values, and hence efficiency, are not observable. These are the two reasons for studying the effect of market selection on market outcomes in a laboratory setting. A common result of existing laboratory studies is that market efficiency improves with the experience of traders (King et. al, 1993; Noussair and Powell, 2010). Specifically, when all traders repeat the same market setting in the same group, markets display fewer bubbles, and bubble episodes decrease with the amount of experience. Further work (Dufwenberg et. al, 2005; Xie and Zhang, 2012) relaxes the condition that all traders have the same level of experience and shows that bubbles in mixed-experience markets are as rare as in fully experienced markets. The Gladyrev: Ural Federal University, Department of Economics; Powell and Shestakova: University of Vienna, Department of Economics. We thank Karl Schlag, Gerhard Sorger, Bluma Powell and seminar participants in Vienna for comments. All remaining errors are our own. 1

2 summary of this line of research is an experience effect: markets populated by more experienced traders are more likely to be efficient. 1 However, all the studies that find the experience effect in mixed-experience markets ignore market selection, as they randomly pick experienced traders to stay in the markets after the inflow of inexperienced traders happens. The main focus of the present study is the sensitivity of mixed-experience markets to the previous success of experienced traders. In our experiment, we create three types of mixed-experience markets that differ in the previous luck-adjusted earnings of experienced traders and observe how these markets differ in their price efficiency. 1 Experimental Design Following Dufwenberg et al. (2005), we first allow a group of six traders to gain experience by participating in three consecutive markets and then use them to form mixed-experience markets that each consists of two experienced and four inexperienced traders. 2 Unlike Dufwenberg et al. (2005), who randomly select some of the experienced traders to form one mixed-experienced market per session, we take all six experienced traders, split them into least, moderately, and most successful pairs based on their luck-adjusted aggregate earnings in the first three markets, and form three mixed-experience markets per session. Hence, our treatment variable is the success level of experienced traders in mixed-experience markets. Traders are always aware of the market composition, but not of the success level of experienced traders including themselves. 3 A parametric structure of all our markets is identical. The six traders trade 18 units of an asset during a period of 20 minutes. At two-minutes intervals, each unit of the asset pays a dividend that is either 0 or 20 cents with equal probability. Since the expected dividend is 10 cents, the risk-neutral fundamental value of the asset after t dividend payments equals fv(t) = (10 t) 10 cents, i.e. 100 cents before the first dividend payment and 0 cents after the last dividend payment. Each trader begins a market with an endowment of three assets and 400 Euro cents. 4 Earnings from a market are equal to final cash holdings, with expected earnings being 700 Euro cents. Trading is done via a computerized continuous double-auction with an open order book that resets after every dividend payment. Borrowing of cash and shares is not allowed. Seven experimental sessions, each consisting of six markets, were conducted at the Vienna Center for Experimental Economics (VCEE) at the University of Vienna. Subjects were recruited using ORSEE (Greiner, 2004). The experiments were computerized using z-tree (Fischbacher, 2007). Each session began with all subjects reading a common set of instructions and answering a set of control questions. Those subjects who were unable to answer all control questions correctly in 15 minutes were asked to leave the experiment. 5 Conditional on at least 18 subjects having answered all control questions correctly, the experiment continued with 6 subjects randomly selected to become experienced and 12 subjects randomly selected to partic- 1 Hussam et. al (2008) argue that the experience effect depends on having a stationary environment. 2 Dufwenberg et al. (2005) consider two treatments, one with one-third and the other with two-thirds of experienced traders in mixed experienced markets. We focus solely on the former scenario. 3 ref:todo: on the effect of revealing rankings 4 The parametric structure of our markets differs from the one used in Dufwenberg et al. (2005) only with regard to original endowments. Half of their traders each started with six assets and 200 cents, and each of the other traders started with two assets and 600 cents. We chose homogeneous endowments because, if anything, this improves subjects comprehension without affecting market efficiency as has been argued in ref:todo. Additionally, original cash-to-asset ratio is higher in our markets (2400 cents to 18 shares as opposed to 2400 cents to 24 shares), which, if anything, stimulates a higher volume of trade (ref:todo). 5 On average, two subjects per session failed this way. One session was canceled for the reason that less than 18 subjects had answered all control questions correctly. 2

3 Figure 1: Time Series of Median Period Prices ipate in a comprehensive risk-preference elicitation experiment in the meantime. Immediately before entering the market for the first time, subjects received an extensive twenty-minutes training for the trading interface and market structure. Instructions, control questions, and screen shots from the training round can be found in the Appendix. 6 The experiment concluded with a brief questionnaire that included three questions of the cognitive reflection test (Frederick, 2005). Each session took two hours and thirty minutes and the average earnings were 32 Euro per subject. 2 Price Efficiency across Markets To fix the notation, we refer to four repetitions of a market within a session as Rounds 1 4, and to three treatments in Round 4 as LOW, MED, and HIGH corresponding to the experienced traders being least, moderately, and most successful in Rounds 1 3 of their session. 2.1 General Price Patterns Figure 1 demonstrates the time series of median period prices from individual markets along with the fundamental values, the round averages for Rounds 1 3 and the treatment averages for Round 4. 7 The fully inexperienced market (Round 1) exhibits a typical bubble-and-crash pattern without a clear evidence of this pattern being mitigated by experience (Rounds 2 and 3). In the mixed-experience market (Round 4), the same pattern prevails, with the highest prices in the LOW-treatment, the lowest prices in the MED-treatment, and the HIGH-treatment being roughly in between. To measure price efficiency of our markets, we use a minor modification of commonly used bubble measures from Haruvy and Noussair (2006). Table 1 shows these measures for every 6 Experimental procedure constitutes yet another difference of our experiment from the one in Dufwenberg et al. (2005). Changes in instructions, exclusion of subjects unable to answer control questions, and provision of an elaborate training round were all aimed at improving subjects comprehension, as confusion and misunderstanding were shown to be important drivers of bubbles (Kirchler and Huber 2012; Kirchler et al. 2012). Additionally, we replaced crossword puzzle solving with risk-preference elicitation task for those subjects who 3

4 Table 1: Observed Values of Bubble Measures Session Round Mean Average Bias LOW MED HIGH Average Dispersion LOW MED HIGH Note: The table reports the observed values of bubble measure for each market along with round and treatment averages. Average Bias = (Pt F V t )/10 and Average Dispersion = P t F V t /10, where P t and F V t are median price and fundamental value in period t. For periods without transactions the average of the latest bid-ask spread is used instead of the average price. market along with the round averages for Rounds 1 3 and the treatment averages for Round 4. Average Bias measures overpricing and equals the average, across all 10 periods in a market, of the per-period deviation of the median price from the fundamental value. Average Dispersion measures mispricing and equals the average, across all 10 periods in a market, of the absolute per-period deviation of the median price from the fundamental value. 8 As follows from Figure 1, median period prices never drop below fundamentals in many of our markets, explaining the same values of Average Bias and Average Dispersion in those markets. In our statistical analysis, we treat one session as one independent observation of each round and treatment outcome. We transform continuous measures into categorical data to run the binomial tests comparing measures between rounds, between treatments, and to benchmarks. Table 2 summarizes the transformed data. With our seven independent observations, the difference is statistically significant at 10%-level if it is of the same sign in six cases, and at 1%-level if it is of the same sign in all seven cases. Observation 1: Bubbles occur in all rounds and treatments. However, bubbles are relatively small in fully experienced markets and mixed-experience markets with moderately successful experienced traders. Support: As follows from Table 1, Average Dispersion is greater than zero in all 42 markets. There are only two occurrences of a negative Average Bias, one in the fully experienced market and one in the mixed-experience market with moderately successful experienced traders, both were chosen to play the role of inexperienced in the last market. 7 For periods without transactions with 11 such periods in the entire experiment the middle of the latest available bid-ask spread is used instead of the median price. 8 These two measures differ from RD and RAD defined in Stock et al. (2010) in two respects. First, they we do not divide our measures by the average market fundamental value, which is equal 55 cents in all our markets, and hence this does not affect our hypothesis testing. Second, median rather than average period prices are used in our measures, which makes them less sensitive to outliers. 4

5 Table 2: Frequencies of Observed Differences Difference Average Bias Average Dispersion Individual Markets R1 > * R2 > * R3 > R4-LOW > 10 6 * 6 * R4-MED > R4-HIGH > 10 6 * 7 *** Experienced vs. Inexperienced Markets R1 > R3 4 4 Mixed-Experience Markets R1 > R4-LOW 2 3 R4-LOW > R3 7 *** 7 *** R1 > R4-MED 5 5 R4-MED < R3 6 * 5 R1 > R4-HIGH 3 4 R4-HIGH > R3 4 4 Treatment Differences R4-LOW > R4-MED 7 *** 6 * R4-HIGH > R4-MED 5 4 R4-HIGH > R4-LOW 3 3 Note: The table shows the number of sessions (out of seven) for which a given difference holds. The difference is statistically significant at 10%-level if it holds in six sessions ( * ), and at 1%-level if it holds in all seven sessions ( *** ). happening in the same session. This evidence allows us to conclude that median period prices are above fundamentals in all rounds and treatments, with the difference being statistically significant. However, Average Dispersion and Average Bias are not significantly greater than the expected dividend of 10 cents in the fully experienced market (4/7 for both measure) and in the mixed-experience market with moderately successful experienced traders (3/7 for Average Bias and 4/7 for Average Dispersion). Additionally, overpricing measured by Average Bias is not significantly greater that the expected dividend of 10 cents in the first two repetitions of the market (5/7 and 4/7 correspondingly), with mispricing measured by Average Dispersion being significantly greater than that in these two rounds. 2.2 Experienced and Inexperienced Markets The experience effect is thought to be one of the most robust findings in experimental asset markets literature (Smith et al. 1988, ref:todo) Up to date, it has only been reported that the experience effect is not present when essential characteristics of the market environment, such as dividend structure or the number of traders, are changed (Hussam et al. 2008, Xie and Zhang 2012) or the market environment itself crucially differs from the commonly used one (Noussair and Powell 2010, Oechssler et al. 2011). Additionally, intensive training is found to serve the same role in mitigating bubbles as repeated participation in the same market environment (Huber and Kirchler 2012, Kirchler et al. 2012, ref:todo). Our data do not support any of these two earlier results. Observation 2: Bubbles are mitigated neither by training nor by experience. 5

6 Support: If training was effective in eliminating bubbles, we would not have observed them in the first round of our experiment. We have, however, already shown that bubbles are present there. To look at the role of experience, we compare Average Bias and Average Dispersion in the fully inexperienced market (Round 1) and the fully experienced market (Round 3) of the same session. In four sessions, both measures are larger in the fully inexperienced market, and in the remaining three sessions the reverse happens. These differences are insufficient to conclude whether price efficiency is higher in fully experienced or fully inexperienced markets. Our failure to observe any experience effect would not be interesting if our changes in the design compared to previous literature were already known to have such an effect, or if one could argue that the setting we study is less realistic than that studied earlier. However, for all of the changes we identify, we argue that they should, if anything, make it more likely to find an experience effect, and increase the realism of the setting. Therefore, we conclude that this instance has important implications for experimental asset markets generally. 2.3 Mixed-Experience Markets The main message of Dufwenberg et al. (2005) is that bubbles are rare in mixed-experience experimental markets, which better represent real asset markets than commonly used homogeneous ones. They arrive at this conclusion by rejecting the hypothesis that fully inexperienced and mixed-experience markets are identical and being unable to reject the hypothesis that fully experienced and mixed-experience markets are identical. If anything, our data suggest the opposite. However, this should be taken with caution given that even our fully experienced markets are indistinguishable from fully inexperienced ones. Observation 3: Mixed-experience markets are more alike fully inexperienced markets than fully experienced markets. Support: In an attempt to replicate findings of Dufwenberg et al. (2005), we compare mixed-experience markets with fully inexperienced and fully experienced markets within a session. None of our three mixed-experienced markets is statistically different from the fully inexperienced markets, regardless of the measure. However, all seven mixed-experience markets with least successful traders exhibit larger mispricing and overpricing than the preceding fully experienced markets. Additionally, six mixed-experience markets with moderately successful traders exhibit lower overpricing than the preceding fully experienced markets. Our main interest lies in studying how price efficiency of markets is affected by previous trading success of experienced traders. Based on the earlier literature, we conjecture that both top and bottom earners may produce lower price efficiency than average traders. In the case of top earners, this could be achieved due to their ability and motivation to speculate (ref:todo), while in the case of bottom earners, this mainly results from their erroneous trading decisions (ref:todo). Observation 4: In mixed-experience markets, past success of experienced matters. Prices exceed fundamentals by a highest margin in markets with least successful experienced traders, followed by markets with most successful experienced traders, and then by markets with moderately successful experienced traders. Support: To test for the treatment effect, we make three pairwise comparisons of three mixed-experience markets within a session. In seven (six) sessions, overpricing (mispricing) is larger when experienced traders are least successful compared to the case when they are moderately successful, making the difference statistically significant. In five (four) sessions, overpricing (mispricing) is larger when experienced traders are most successful compared to the case when they are moderately successful. At the same time, we observe three sessions with both overpricing and mispricing being larger when experienced traders are most successful compared to the case when they are least successful. This makes mixed-experience markets 6

7 Figure 2: Bubble Measures across Success Levels with most successful experienced traders indistinguishable from the other two mixed-experience markets. Combined with the fact that least successful experienced traders when mixed with inexperienced traders produce significantly lower price efficiency than moderately successful ones, we conclude that market efficiency in mixed-experience markets with most successful experienced traders is roughly in between the others two. Additional support for the success effect comes from plotting bubble measures of mixedexperienced markets against previous luck-adjusted earnings of experienced traders (Figure 2), which have been used to split them into three success treatments. The corresponding regression on the data clustered at the session level finds a statistically significant quadratic relationship between the previous success of experienced traders and price efficiency of mixed-experience markets (p < 0.01): markets populated with more extreme earners have higher bubbles. The strong relationship between the previous earnings of experienced traders and the price efficiency of mixed-experienced markets should be treated with caution though. Larger dispersion in earnings across success levels within a session is likely to indicate larger bubbles in the earlier rounds of this session (Hirota and Sunder 2007), which are themselves endogenous with respect to the traders behavior. The effects of traders innate behavior and the mere fact of previously experiencing larger bubbles caused by other traders cannot be disentangled without an additional experiment designed specifically to address this issue. 3 Conclusion This study examines the sensitivity of mixed-experience markets to the previous success of experienced traders. The results suggest that markets populated by experienced traders with more extreme previous earnings exhibit significantly lower price efficiency than those with more modest earnings. We expect real world markets to display exactly these kinds of characteristics, and thus our results call into question the external validity of previous results suggesting that experience always has a dampening effect on bubble size. Earnings are shown to be contemporaneously correlated and predict several types of individual behavior. We identify strong patterns that show that more extreme earners behave differently from their peers. More extreme earnings are associated with higher trading activity and riskier portfolios, whereas higher transaction risk is associated with lower earnings. In terms of predicting future behavior, current earnings are shown to forecast aspects of trading activity, portfolio risk and transaction risk. Finally, while not a central focus of our study, we fail to consistently replicate an important result regarding the presence of an experience effect in a repeated market setting. Due to the 7

8 limited dataset of this study (7 sessions), we can only point out that this calls into question the consistency of this effect. More generally, the summary of this analysis is that earnings are a key indicator of both market and individual behavior. If, as we have argued above, the distribution of previous earnings in real markets is not uniform, then the findings reported here have important implications for the analysis of experimental markets. We suggest that additional research is merited on experience, earnings, and the interplay between the two in market settings. References Alchian, Uncertainty, evolution, and economic theory. The Journal of Political Economy, Dufwenberg, Lindqvist, and Moore, Bubbles and Experience: An Experiment. American Economic Review, 95(5): Fischbacher, z-tree: Zurich toolbox for ready-made economic experiments. Experimental Economics, 10(2): Greiner, An online recruitment system for economic experiments. Forschung und wissenschaftliches Rechnen Vol. 63, Fama, Efficient capital markets: A review of theory and empirical work. Journal of Finance, 25(2), Haruvy, Lahav, and Noussair, Traders expectations in asset markets: Experimental evidence. American Economic Review, 97(5), Haruvy and Noussair, The effect of short selling on bubbles and crashes in experimental spot asset markets. Journal of Finance, 61(3), Hirota and Sunder, Price bubbles sans dividend anchors: Evidence from laboratory stock markets. Journal of Economic Dynamics and Control, 31(6): Hussam, Porter and Smith, Thar She Blows: Can Bubbles Be Rekindled with Experienced Subjects?. American Economic Review, 98(3): Kirchler and Huber, The impact of instructions and procedure on reducing confusion and bubbles in experimental asset markets. Experimental Economics, 15(1): Kirchler, Huber, and Stckl, Thar She Bursts: Reducing Confusion Reduces Bubbles. American Economic Review, 102(2): Lim and Brooks, The evolution of stock market efficiency over time: a survey of the empirical literature. Journal of Economic Surveys, 25(1), Noussair and Powell, Peaks and Valleys: Price Discovery in Experimental Asset Markets with Non-monotonic fundamentals. Journal of Economic Studies, 37(2), Smith, King, Williams, and Van Boening, The Robustness of Bubbles and Crashes in Experimental Stock Markets. Non Linear Dynamics and Evolutionary Economics. Ed. Day and Chen. Oxford Press Stckl, Huber and Kirchler, Bubble measures in experimental asset markets. Experimental Economics, 13(3), Xie and Zhang, Bubbles and Experience: An Experiment with a Steady Inflow of New Traders. Working paper. 8

9 Appendix 3.1 Screen shots from the experiment Figure 3: Instructions and control questions screens 9

10 Figure 4: Training screen: explaining how to make bids 10

11 Figure 5: Pre-trading and trading screens 11