Model Risk in Portfolio Optimization: A Multi-Agent Framework for Dynamic Portfolio Optimization

ანოტაცია

Financial markets are complex, high-dimensional, and constantly evolving, posing significant challenges to traditional portfolio optimization methods. Classical mean–variance frameworks often rely on restrictive assumptions (such as Gaussian returns and static parameters estimated from historical data) that can lead to unstable and suboptimal portfolios when markets exhibit non-linear dependencies, heavy tails, or regime shifts. In this paper, we propose a novel multi-agent simulation and model selection system (MAS2) for dynamic portfolio optimization that addresses two key issues: (i) robust parameter estimation under limited, noisy data, and (ii) adaptive model selection in non-stationary market conditions. In our framework, multiple agents, each based on a distinct stochastic model class, collaboratively learn and adapt by generating synthetic market data, updating their model parameters via Bayesian inference, and undergoing rigorous performance evaluation using Bayesian model selection criteria. Poorly performing models are iteratively pruned, and the process continues until convergence to a stable set of models and portfolio strategies. The resulting portfolio allocations inherently account for model uncertainty and adapt to changing market dynamics, offering a robust alternative to static single-model approaches. We present the framework and discuss its properties; empirical assessment is left for future work.