A1474
Title: AI-driven residual learning with robust factor models for stock return prediction
Authors: Zhuolan Liu - University of Macau, FBA (China) [presenting]
Xinyi Ren - University of Macau (China)
Lianjie Shu - University of Macau (China)
Abstract: Accurate stock return prediction remains a central challenge in financial econometrics, particularly under structural instability and nonlinear market dynamics. A hybrid framework is proposed that integrates robust factor modeling with machine learning to improve predictive performance. Specifically, Huber Principal Component Analysis (HPCA) is employed to extract stable latent factors from high-dimensional financial data, followed by vector autoregression (VAR) to capture temporal dynamics. A residual neural network is then applied to model nonlinear components unexplained by traditional factor structures. Using U.S. equity data from 1973 to 2019, results indicate that the proposed approach improves out-of-sample predictive accuracy compared to standard benchmarks such as CAPM and Fama-French models, as well as standalone machine learning methods. The findings highlight the value of combining structured econometric models with AI-driven techniques, providing a robust and flexible framework for financial forecasting.
