Machine Learning for Financial Risk Management: Modeling Time-Varying Factor Sensitivities Using Factor Variational Autoencoders

Abstract

Accurate measurement of time-varying systematic risk exposures is essential for robust financial risk management. Conventional asset pricing models, such as the Fama-French three-factor framework, assume constant factor loadings, limiting their ability to capture shifts in risk during volatile market conditions. This study employs a Factor Variational Autoencoder (FactorVAE) to model nonlinear, dynamic sensitivities to the size (SMB) and value (HML) factors using daily returns for 100 S&P 500 constituents from January 2018 to December 2024. The model extracts ten statistically independent latent risk factors, reducing reconstruction error by 44% relative to rolling-window Ordinary Least Squares (OLS). These latent factors align with economically interpretable structures—volatility regimes, sector-specific risk dynamics, and macroeconomic cycles—providing enhanced insight into evolving firm-level exposures. The dataset exhibits pronounced fat tails, validating the use of nonlinear models for improved tail-risk estimation. Factor trajectories reveal continuous evolution in exposures, with earlier detection of structural changes during the COVID-19 recovery and Federal Reserve tightening cycles compared to linear benchmarks. Principal Component Analysis of the latent space identifies distinct market regimes and gradual transitions, facilitating forward-looking regime classification. The results demonstrate that machine learning-based dynamic factor modeling can materially improve systematic risk measurement, enable proactive regime monitoring, and support more responsive hedging and capital allocation strategies.