Mesoscale impact of trader psychology on stock markets: a multi-agent AI approach

Johann Lussange, Stefano Palminteri, Sacha Bourgeois‐Gironde, Boris GutkinarXiv 2019 · DOI · arXiv

Abstract

Recent advances in the fields of machine learning and neurofinance have yielded new exciting research perspectives in practical inference of behavioural economy in financial markets and microstructure study. We here present the latest results from a recently published stock market simulator built around a multi-agent system architecture, in which each agent is an autonomous investor trading stocks by reinforcement learning (RL) via a centralised double-auction limit order book. The RL framework allows for the implementation of specific behavioural and cognitive traits known to trader psychology, and thus to study the impact of these traits on the whole stock market at the mesoscale. More precisely, we narrowed our agent design to three such psychological biases known to have a direct correspondence with RL theory, namely delay discounting, greed, and fear. We compared ensuing simulated data to real stock market data over the past decade or so, and find that market stability benefits from larger populations of agents prone to delay discounting and most astonishingly, to greed.

Introduction

Financial markets are complex "bottom-up" systems where global phenomena emerge from the collective interactions of individual actors. While traditional econometrics often treats market movements as abstract statistical events, the field of neurofinance seeks to understand the biological and psychological drivers behind these decisions. However, a significant gap remains at the mesoscale: how do specific individual cognitive biases aggregate into large-scale "stylized facts" such as volatility clustering or market crashes? This research utilizes a high-fidelity Agent-Based Model (ABM) to bridge this gap, simulating thousands of interacting traders to observe how psychological "infections" alter market-wide stability and liquidity.

Methodology

The study employs a Multi-Agent System (MAS) where 500 autonomous agents trade via a centralized double-auction limit order book. Each agent's behavior is governed by two Reinforcement Learning (RL) modules: a Forecasting module that predicts price movements using chartist and fundamentalist inputs, and a Trading module that manages portfolio allocation.

To explore the impact of psychology, the researchers mathematically mapped three core human biases into the RL framework:

  1. Delay Discounting: Shortening the investment horizon, prioritizing immediate rewards.
  2. Fear: A heuristic triggering "sell" orders during high volatility or low liquidity, independent of price direction.
  3. Greed: A heuristic triggering "buy" orders upon detecting positive trends, ignoring risk indicators.

The simulator's performance was validated against a decade of real-world data (2007–2018) from the London Stock Exchange (LSE), covering 640 stocks.

Results and Discussion

Liquidity and Short-termism

One of the most significant findings involves the role of "delay discounting." Contrary to the common criticism of short-termism in financial policy, the simulation suggests that agents with shorter investment horizons are vital for market liquidity. As the population of these agents increases, trading volumes rise and the bid-ask spread—a key measure of transaction costs and market efficiency—steadily decreases.

Figure 4: Mean bid-ask spread (as % of price) decreases as the population of short-term/biased agents increases, indicating improved market liquidity.

The Stability Paradox: Greed vs. Fear

The study reveals a counter-intuitive relationship between bias and stability. While "fearful" agents create feedback loops of illiquidity and panic-selling, "greedy" agents act as a stabilizing force. By consistently buying during uptrends and ignoring volatility signals, greedy agents provide a "demand floor" that prevents the price spirals characteristic of market crashes.

Conversely, the "fear" bias is the primary driver of systemic instability. As shown in the simulation results, increasing the prevalence of fear-driven heuristics leads to a marked increase in the frequency of market crashes.

Figure 5: The number of market crashes increases significantly as the prevalence of fear-based psychological traits grows within the agent population.

Conclusion

By integrating neurobiological insights into a multi-agent RL framework, this research demonstrates that individual psychological biases have profound and often non-obvious impacts on global market dynamics. While fear remains a potent contagion for crashes, traits often viewed negatively—such as greed and short-termism—appear to play essential roles in maintaining market liquidity and preventing panic-driven collapses. Future work will expand this framework to include short-selling, a wider array of cognitive biases, and cross-asset systemic risks.

Summary updated

Abstract

Recent advances in the fields of machine learning and neurofinance have yielded new exciting research perspectives in practical inference of behavioural economy in financial markets and microstructure study. We here present the latest results from a recently published stock market simulator built around a multi-agent system architecture, in which each agent is an autonomous investor trading stocks by reinforcement learning (RL) via a centralised double-auction limit order book. The RL framework allows for the implementation of specific behavioural and cognitive traits known to trader psychology, and thus to study the impact of these traits on the whole stock market at the mesoscale. More precisely, we narrowed our agent design to three such psychological biases known to have a direct correspondence with RL theory, namely delay discounting, greed, and fear. We compared ensuing simulated data to real stock market data over the past decade or so, and find that market stability benefits from larger populations of agents prone to delay discounting and most astonishingly, to greed.
BibTeX
@misc{lussange2019mesoscale,
  author = {Johann Lussange and Stefano Palminteri and Sacha Bourgeois‐Gironde and Boris Gutkin},
  title = {{Mesoscale impact of trader psychology on stock markets: a multi-agent AI approach}},
  year = {2019},
  eprint = {1910.10099},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1910.10099},
}
Figure