Self-supervised risk factor model using dual Recurrent State Space Models

Asset pricing models are crucial in finance for asset valuation, portfolio management, and investment strategy formulation. The integration of deep learning into these models leverages their ability to distill meaningful patterns from extensive datasets, recognize complex nonlinear interdependencies, and improve predictive precision. However, the inherently noisy nature of financial data poses significant challenges in model training and generalizability to new data. Addressing these challenges, our paper introduces an innovative risk factor model framework, grounded in a Recurrent State Space Model (RSSM). We have designed a dual RSSM structure that separately models the overall market and individual assets. This approach facilitates a deep comprehension of market trends and captures the temporal dynamics specific to each asset. Furthermore, we present a novel self-supervised technique to enhance our model's performance. Our model provided a better understanding of volatility in the U.S. stock market and achieved a 41% improvement in predicting asset returns compared to the best-performing model among previous studies. We also demonstrated the ability to generate a 79% increase in risk-adjusted returns through investment simulations. Our code is available at https://github.com/Osj1614/dualrssm.