Factor Structure of Cross-Sectional Stock Returns : A Machine Learning Approach

We propose a unsupervised learning approach to construct latent factor model for cross sectional asset returns where firm characteristics instrument for the dynamic factor exposures. Firm characteristics are clustered with consideration to their prior economic content. Our method can also be viewed as a constrained version of Instrumented Principal Components Analysis (IPCA) model by Kelly et al (2019). We apply the method to a set of 94 firm characteristics to form clustered factor models. The model outperforms all benchmark models, and outperforms IPCA latent factor model out-of-sample