Overconfidence—the tendency of investors to overestimate the precision of their information or forecasting ability—is a widely studied behavioral bias with profound implications for financial markets. It drives excessive trading, underestimation of risk, and persistent mispricing, phenomena that traditional rational expectations models like CAPM or REE struggle to explain. While early behavioral models introduced boundedly rational agents like noise or momentum traders, they lacked a systematic way to capture belief heterogeneity and interaction. This paper addresses that gap by proposing a Mean Field Game (MFG) framework in which a continuum of overconfident agents interact through endogenous price feedback. Agents update their beliefs using Kalman or Extended Kalman Filters and make trading decisions based on biased posteriors, with prices determined via a demand-aggregating price impact function. This approach enables rigorous analysis of how overconfidence shapes market volatility, mispricing, and deviations from informational efficiency. We build on prior behavioral and heterogeneous agent models to investigate how varying degrees of overconfidence affect asset pricing dynamics across different market regimes.
Mean Field Games (MFGs) are a mathematical modeling framework used to analyze decision-making in systems with a large number of interacting agents, where each agent is individually insignificant but collectively influences the environment. In this research, MFGs are employed to model financial markets populated by a continuum of overconfident investors—agents who systematically overestimate the precision of their private information. The MFG framework enables the study of belief heterogeneity, learning dynamics, and endogenous price formation by treating each agent’s optimal strategy as a response to the aggregate behavior of the population. Through both static and dynamic formulations, including Hamilton-Jacobi-Bellman and Fokker-Planck equations, the model captures how overconfidence amplifies mispricing and volatility. Agents update their beliefs using Kalman or Extended Kalman filters and determine asset demand based on biased posterior estimates. The equilibrium emerges as a feedback loop where collective beliefs shape market prices, which in turn influence future beliefs and strategies. This approach allows for tractable analysis of complex market phenomena like sentiment cascades, informational inefficiencies, and heterogeneous reactions between institutional and retail investors.
The first simulation in this research models a financial market using the dynamic Mean Field Game (MFG) framework to study how overconfidence among investors affects market efficiency, price volatility, and mispricing. It simulates a population of agents who update beliefs about an unobservable fundamental value using Kalman filtering and form trading strategies based on these beliefs. The agents assume Constant Absolute Risk Aversion (CARA) utility and determine optimal asset demand by maximizing expected utility given perceived mispricing. Overconfidence is modeled by scaling the precision of private signals with a parameter k>1, leading agents to overweight their private information in forming posterior beliefs. The simulation is conducted under four distinct market scenarios—bullish and bearish regimes with and without overconfidence—by varying the fundamental value's drift and the investor confidence level. Market prices aggregate individual demands through a price impact function, completing the feedback loop between beliefs, trading, and price formation.
Numerically, the simulation solves a system of coupled partial differential equations: the Hamilton-Jacobi-Bellman (HJB) equation for the value function and the Fokker-Planck-Kolmogorov (FPK) equation for the belief distribution. The HJB is solved backward in time using an implicit Euler method and central finite differences, while the FPK is solved forward in time on the same discretized grid. Neumann boundary conditions ensure that mass and value gradients do not escape the belief space. The simulation tracks belief evolution, mispricing, and volatility across 100 runs for each scenario, averaging results to reduce stochastic noise. Key findings reveal that higher overconfidence increases volatility and average mispricing, destabilizing prices, especially in trending regimes. However, the model's limitations include simplifications such as fixed agent preferences, a linear pricing rule, and constant volatility, all of which constrain realism. Furthermore, the simulation does not yet distinguish between investor types or account for nonlinear signal processing elements that are later addressed in the enhanced model.
The second simulation builds upon the initial Mean Field Game (MFG) framework by introducing key enhancements to improve the realism, granularity, and interpretive power of the model. It transitions from solving partial differential equations to an agent-based structure, where each investor—retail or institutional—maintains a personalized belief, uncertainty estimate, and dynamically evolving level of overconfidence. This new framework introduces Extended Kalman Filtering (EKF) to model nonlinear signal observations, particularly for retail agents, who receive distorted or sentiment-driven signals. Agents update their beliefs using the appropriate filter and adjust trading behavior accordingly, with trading aggressiveness scaled by their confidence k_{i,t}. These changes allow the model to account for informational asymmetries, diverse learning rates, and sentiment cascades that emerge from herding behavior. The market price is then determined endogenously by aggregating demand across all agents, incorporating behavioral feedback into price evolution. This approach enables simulations of more complex market regimes—bullish, bearish, steady, and volatile—using a stochastic drift model for fundamentals, providing richer insights into how overconfidence interacts with uncertainty and trend direction.
Compared to the first simulation, these enhancements are critical for capturing empirical features of real-world markets. The agent-based formulation introduces heterogeneity in both belief formation and trading strategy, allowing for endogenous clustering of behavior, the emergence of fat-tailed price distributions, and temporally correlated volatility—features commonly observed in financial markets but absent in the earlier PDE-based model. By including confidence bands derived from Kalman posterior variances, the model also visualizes belief uncertainty and how it evolves under different information environments. Most notably, the steady market scenario demonstrates that mispricing can worsen over time purely due to informational ambiguity, rather than volatility or trend—a key insight that highlights the destabilizing role of uncertainty itself. From here, future work will focus on expanding the model’s empirical relevance. This includes calibrating the simulation to real financial data, particularly in the context of speculative bubbles and crash dynamics; testing regime-switching behavior using historical examples such as the 2008 financial crisis or 2020 pandemic onset; and comparing model-generated uncertainty with recent market responses to geopolitical developments like tariff policies. By grounding the model in observable market behavior, it will be possible to assess the predictive validity of overconfidence dynamics and evaluate policy implications for financial stability.
This research has significant applications in both academic and practical contexts, particularly within behavioral economics and financial market analysis. By formalizing the effects of overconfidence through a rigorous Mean Field Game framework, the model provides a quantitative tool for understanding how belief distortions propagate through market systems, leading to mispricing, excess volatility, and feedback-driven instabilities. Such insights are valuable for regulators aiming to monitor systemic risk, for asset managers seeking to anticipate sentiment-driven market movements, and for policy makers interested in designing interventions that mitigate the effects of cognitive biases. In the broader field of behavioral economics, this research contributes a dynamic, population-level perspective that moves beyond representative-agent models, capturing heterogeneity in belief formation and the endogenous evolution of confidence. It supports the growing recognition that bounded rationality, informational asymmetries, and psychological reinforcement mechanisms are not peripheral anomalies but fundamental drivers of aggregate outcomes in complex systems like financial markets.