Is Rational Asset Valuation in Unstable Financial Environments Possible? Experimental Evidence Based on the Bandit Problem∗

An important issue in financial decision-making is the way people process new information. Prior studies have questioned the ability of people to use Bayes ’ law in decision-making. None of those studies however probe situations similar to those typically encountered in financial markets. Here we explore, both theoretically and empirically, whether agents can apply Bayes ’ law in a finance environment, cap-tured as a nonstationary bandit task. In it, we isolated the instability encountered in modern financial markets in the form of sudden changes (“jumps”) in the return processes. From subjects ’ choices, we determined whether their learning in the task reflected optimal Bayesian inference instead of simple “Reinforcement Learning. ” In contrast to the latter win-keep lose-switch heuristic, the Bayesian models accommo-date nonstationarity either by tracking the probability of a jump or by dynamically adjusting the memory of learning through jump detection. Both Bayesian models fitted our data better than the Reinforcement Learning model did. This result sug-gests that agents may be better fit to learn in financial markets than previously thought. More broadly, it suggests that humans are wired to face very complex set-tings, and should prompt a reexamination of the scope of the bounded rationality paradigm