Finance - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemHuman, Artificial Intelligence and Tail Risk( 2022)I study learning efficiency of artificial and human agents in the environment of financial markets. One prominent feature that distinguishes this environment is tail risk, which means that outliers are more frequent and substantial relative to Gaussian outliers. Failure to account for tail risk deteriorates learning efficiency, causing agents to derail from optimal actions. In the dissertation, I explore improvements to learning by artificial agents under tail risk, and whether human learning exhibits similar improvements. Finally I study to what extent agents' interactions and intelligence level would cause or amplify tail risk. A key success of artificial intelligence has been reinforcement learning. I first show that even the most advanced reinforcement learning protocol yields sub-optimal behavior in an environment with tail risk. Inspired by the concept of statistical efficiency, I propose a solution that nicely complements a recent protocol -- distributional reinforcement learning -- and improves the performance of algorithms. I show that the proposed algorithm learns much faster and is robust once it settles on a policy. Thus, efficiency gains are possible for artificial agents. Do humans exhibit the same kind of adjustment in an environment of tail risk? In the second study, I design an experiment to examine whether and how efficiency concerns drive human learning of stochastic rewards. While I find substantial heterogeneity, overall the answer is affirmative. Efficiency gains translate into enhanced choice confidence, except when participants fail to discover the most efficient estimator. In finance, the real causes of tail risk remain elusive. One conjecture is that, even without triggers from any extreme event, tail risk emerges because of agents' interactions in the marketplace. Motivated by the zero-intelligence and machine learning literature, I propose a paradigm to approach this conjecture in the third study. The paradigm comprises a single-widget economy, a continuous open-book market, and a group of trading agents with different intelligence levels. I demonstrate that trading generates excessive tail risk even when the underlying economic shifts follow a Gaussian law. Introducing a profit-seeking market maker further increases leptokurtosis, but the tail risk is not worsened. The latter suggests that tail risk and leptokurtosis may need to be distinguished.