Market behavior from traders' cognitive traits: a multi-agent human reinforcement learning approach

Co-directors:
Boris Gutkin
Damien Challet
Stefano Palminteri

Jury:
Iryna Veryzhenko - Rapporteuse
Angela Yu - Rapporteuse
Michael Benzaquen - Président du jury
Laura Fontanesi - Examinatrice

Abstract:
This PhD thesis investigates the collective impact of individual decision-making on the stock market. Traditionally, mainstream economists assume traders to be rational individuals capable of exploiting market inefficiencies. However, individuals can be irrational. One crucial bias is confirmation bias, which pushes individuals to accept information that confirms their own choices, while disregarding evidence that contradicts them. Since the market aggregates information across traders, we ask whether it mitigates or accumulates the macroscopic effects of individual biases.
Inspired by previous work, we develop a multi-agent system simulating the stock market, where prices are entirely driven by agents' actions. Our major modelling contribution is the design of agents based on a reinforcement learning model that best captures human learning and decision-making patterns in some well-established behavioral economics experiments. We validate our market model both at macroscopic and microscopic levels. The simulated market alternates between market efficiency regimes and speculative bubbles, displaying typical statistical properties of real-world markets, such as negative skewness, excess kurtosis, and excess volatility. At the individual level, agents that are either more impatient or less informed develop a tendency to speculate over market trends; if more patient or more informed, they favour fundamentalism. The main findings on the impact of confirmation bias also align with our expectations both macroscopically and microscopically. Choice inertia from confirmation bias promotes market predictability and deepens polarisation of strategy preferences among agents. Additionally, confirmation bias is found to be an endogenous source of negative asymmetry in price returns.
Overall, this work provides two key contributions. Firstly, it demonstrates the potential of leveraging experimentally-validated cognitive models to disentangle the endogenous and exogenous factors that determine both individual and collective system properties. Secondly, it offers a deeper understanding of how individual biases can influence financial markets.