Navigating Uncertainty with Bayesian-Enhanced Multi-Factor Models in Adaptive Asset Pricing

Weaknesses of Traditional Multi-Factor Models

Traditional multi-factor models often lack the ability to adapt to market regime changes and efficiently handle uncertainty. For that reason, our co-founder and head of data analytics, Michal Dufek, led research on a sophisticated framework incorporating Bayesian feature selection, penalized linear models, Bayesian neural networks, and Dynamic Model Averaging. This work fits into the broader context of Stochastic Optimal Control in Asset Pricing, which explores various methods for cross-sectional asset pricing under uncertainty, such as Bayesian approaches and advanced machine learning techniques like deep reinforcement learning using hierarchical graphs and graph attention networks. The common thread linking these methodologies is the focus on handling uncertainty and dynamically adapting to market changes, pivotal in stochastic control problems.

And what did we come up with?

Bayesian-Enhanced Multi-Factor Models in Adaptive Asset Pricing Results

Initial experiments indicate the Bayesian feature selection successfully identifies a reduced set of impactful factors. Subsequent models, both penalized linear and Bayesian neural networks, show improved prediction accuracy and out-of-sample robustness. Furthermore, the DMA layer shows a considerable improvement in adapting to market regime changes, as evidenced by various financial metrics like the Sharpe ratio and maximum drawdown.

Want to learn more? Contact us or visit the 27th European Scientific Conference of Doctoral Students PEFnet 2023 conference where Michal is presenting his results, detailed calculations, and conclusions. As a teaser see the PEFnet paper.