Asset Risk Pricing with Causal Representation Learning and Generative AI: Evidence from Complex Market Environments

Abstract

This study examines asset pricing in high-dimensional financial environments, where traditional factor models face limitations in interpretability and stability, and machine learning approaches remain largely correlation-driven. Existing methods often struggle to identify economically meaningful risk factors and exhibit limited robustness across changing market conditions. To address these limitations, this paper proposes a unified framework that integrates generative modeling with causal representation learning. A generative model is first employed to extract latent structures from high-dimensional financial data, and a causal learning module is subsequently used to identify invariant factors based on structural relationships. These factors are then incorporated into a standard asset pricing model using the Fama–MacBeth regression approach. Empirical results show that the proposed model achieves a cross-sectional R2 of 0.38, compared to 0.26 for the Fama–French five-factor model and 0.31 for a deep learning benchmark. Pricing errors are reduced from 0.054 (FF5) and 0.049 (deep learning) to 0.036. In addition, the model maintains relatively stable performance across different market conditions, including bull, bear, and crisis periods, while benchmark models exhibit a more pronounced decline in explanatory power. These findings suggest that incorporating causal representation improves both explanatory power and robustness in asset pricing. By linking latent structure learning with causal invariance, the proposed framework provides a structured interpretation of risk premia and offers a more stable foundation for asset pricing in complex market environments.