FactorVAE for Probabilistic Cross-Sectional Stock Return Forecasting

Summary

FactorVAE is presented as a probabilistic dynamic factor model that uses a variational autoencoder to extract latent factors from noisy market observations and forecast cross-sectional stock returns. During training, an encoder and decoder use realized future returns to derive useful posterior factors; a separate predictor then learns to estimate those factors from historical observations. At prediction time, only the predictor and decoder are used, which is intended to avoid future information leakage. Treating factors as random variables also allows the model to represent uncertainty and estimate risk alongside expected returns.

The article reports experiments on daily Chinese A-share data from 2010 through 2020, using 20 price and volume features, and compares the approach with linear factor models and several machine-learning baselines. It says FactorVAE outperformed the compared models in cross-sectional prediction and that removing the posterior-guidance method weakened results. The account omits detailed result tables and strategy-performance figures, so the strength and practical significance of the findings cannot be assessed here. Results are tied to the described dataset and setup.

Key ideas

• FactorVAE models latent factors as random variables to represent noisy market data and return uncertainty.
• Training uses future returns to infer posterior factors, while a history-only predictor learns to approximate them.
• Forecasting uses the predictor and decoder to reduce information leakage from future returns.
• The reported evaluation uses Chinese A-share daily data and compares FactorVAE with linear and machine-learning baselines.
• The source reports better prediction results but omits detailed metrics and strategy-performance evidence.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.