A1187
Title: TransDFM: A Bayesian transformer-dynamic factor model for stock market analysis
Authors: Dominic Dayta - Nara Institute of Science and Technology (Japan) [presenting]
Kazushi Ikeda - Nara Institute of Science and Technology (Japan)
Takatomi Kubo - Nara Institute of Science and Technology (Japan)
Abstract: While traditional dynamic factor models (DFM) provide an interpretable framework for decomposing systematic and idiosyncratic behavior in complex systems such as the stock market, they remain restricted by simplifying assumptions (e.g., linearity). Meanwhile, transformer architectures excel at capturing highly non-linear dynamics but often function as "black boxes". The attempt is to combine these two paradigms by introducing the transformer-dynamic factor model (TransDFM), a novel Bayesian framework for financial time series. The DFMs linear factor loading is replaced with a flexible Bayesian neural network embedding and model the latent factor dynamics using a Bayesian structural aligned mixture of VAR (SaMoVAR) process. This leverages the SaMoVAR's structural equivalence to a vector autoregression to enable a principled application of priors over its dynamic attention weights. Applied to the Philippine Stock Exchange (PSE) and S\&P 500, the TransDFM demonstrates strong out-of-sample predictive accuracy and generates well-calibrated uncertainty intervals. Furthermore, despite its highly non-linear structure, the model retains economic interpretability: its primary learned factor shows a statistically significant alignment with systematic market risk as measured by the CAPM beta. The TransDFM provides a powerful new tool that combines the predictive performance of deep learning with the structural analysis of classical econometrics.
