Graph-based multi-factor asset pricing model

Abstract

We propose a latent multi-factor asset pricing model that estimates risk exposure based on firm characteristics and connectivity between assets. To handle connected high-dimensional characteristics, we adopted a graph convolutional network while estimating the connectivity between assets from the correlation of asset returns. Unlike recent literature involving the deep-learning-based latent factor model, we propose a forward stagewise additive factor modeling architecture that constructs latent factors sequentially to maintain the previous stage’s factors. Our empirical results on individual U.S. equities show that the proposed graph factor model outperforms other benchmark models in terms of explanatory power and the Sharpe ratio of the factor tangency portfolio.