FactorVAE for Probabilistic Cross-Sectional Stock Return Forecasting

Summary

FactorVAE is described as a probabilistic dynamic factor model that applies a variational autoencoder to noisy market data to forecast cross-sectional stock returns. In training, an encoder-decoder uses future returns to identify posterior factors that reconstruct returns. A predictor then learns to approximate those factors using historical observations alone. At forecast time, the model uses the predictor and decoder, a design intended to prevent future information from entering predictions. Random latent factors also support probabilistic return and risk estimates.

The article summarizes an evaluation on Chinese A-share daily data covering 2010 to 2020, with 20 price-volume features and comparisons against linear factor models and machine-learning models. It reports that FactorVAE outperformed those baselines and that a variant without posterior-to-prior guidance struggled to learn effective factors. The text provides no detailed numerical metrics or complete strategy-performance results, limiting independent assessment. Its reported evidence is specific to the study's sample and experimental setup.

Key ideas

• FactorVAE treats investment factors as latent random variables to model noisy observations and return uncertainty.
• A posterior factor extractor learns from realized returns, and a predictor approximates those factors from historical data.
• The forecast stage uses only the predictor and decoder to avoid future information leakage.
• The described study evaluates the model on Chinese A-share data against linear and machine-learning baselines.
• Reported improvements lack detailed metrics in this summary and may not generalize beyond the tested setup.

Tags

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