Learning to be smart: can humans learn to improve profitability and risk control in financial trading?

Recently, the efficient market hypothesis has faced strong challenges from various fields, and the purpose of this thesis is to provide empirical evidence for the challenges to the efficient market hypothesis from two perspectives. The first one is from the field of machine learning. While an increasing number of machine learning studies report the high accuracy of stock market prediction, this is not consistent with the efficient market hypothesis which suggests that current stock prices discount available information and that it is not possible to obtain systematic returns by exploiting any predictability of prices. As most of the machine learning studies choose relatively simple test settings, I suspect that the reported high accuracy might result from biased performance measurement. That is, the selection of methodological factors is influential on prediction performance in stock markets. To test my conjecture, I run the benchmark with a comprehensive combination of the methodological factors to collect the performance measures under various settings. Next, I analyze the relationship between the prediction performance and the methodological factors. I find the significant influence of the selection of methodological factor on prediction performance, which means that the reported high prediction performance might be biased and my results are not against the prediction of the efficient market hypothesis.The second challenge is that there is increasing evidence of anomalies in financial markets. This suggests that the underlying rationality principle of the efficient market hypothesis may be flawed. The manner in which individuals learn from experience also remains a matter of debate. The rationality assumption would be justified if individuals follow Bayesian learning, i.e., individuals learn from experience to appropriately adjust their probability estimates and finally make rational and appropriate decisions. To examine the relationship between experience and performance measures, I use linear mixed models to analyze spread trading data. I find that, as individuals gain experience, they increase their degree of risk-taking and realize higher returns. However, thesereturns are subject to greater volatility and, as a result, they achieve lower risk-adjusted returns. Since the individuals following Bayesian learning should be able to appropriately update probability estimates conditioned on new information, their decision choices and their risk-adjusted performance should be improved. My results show that individuals fail to follow Bayesian learning. On the other hand, my results can be explained by reinforcement learning, wherein individuals repeat behavior that was rewarding in the past. Traders may try several trading strategies with different levels of risk. Since higher risk generally brings both higher profits and greater losses, traders who undertake riskier strategies will either make higher profits or suffer greater losses. Those traders making a higher profit are reinforced by the riskier strategies and overlook the underlying risk, which leads to lower risk-adjusted performance. Hence, my results cast doubt on the validity of the rationality assumption.To further explore the degree of rationality with trading data, I propose a method to estimate the degree to which an individual behave like a rational agent, and other behavioral characteristics. The experience weighted attraction (EWA) can be used to estimate the degree of rationality in psychological experiments, but cannot be used with trading data. The reason is that the number of strategies available to decision makers was limited in psychological experiments, but in real-world trading environments, traders have no limits in terms of the strategies they can adopt. We propose a decision-based strategy mapping framework (DSM) to resolve this problem. The DSM is designed to artificially limit the strategy space associated with real-world trading data, by using scenarios. In each scenario, individuals are assumed to have only one decision to make. This allows us to estimate, using data associated with an individual’s real-world trading, their behavioral characteristics associated with EWA. Subsequently, we examine the relationship between the estimated behavioral characteristics of traders and their trading behavior and performance. My results suggest that those traders who behave like rational agents tend to trade more actively. However, surprisingly, those traders who are more rational do not achieve superior trading performance.In conclusion, the findings of this thesis support the efficient market hypothesis that the markets are efficient, at least to the extent to which excess returns cannot be earned with state of the art machine learning techniques. However, the results of learning behavior from individual-level analysis suggest that the rational agent assumption of the efficient market hypothesis is likely to over-simplify individual behavior in the real world.