Deep Learning in Characteristics-Sorted Factor Models

Many view deep learning as a "black box" used only for forecasting. However, this paper provides an alternative application by constructing a structural deep neural network to generate latent factors in asset pricing. The conventional approach of sorting firm characteristics to generate long-short factor portfolio weights is underappreciated nonlinear modeling. First, we describe the complete mechanism for fitting crosssectional returns by firm characteristics through risk factors. Second, unlike statistical models, our model has an economic-guided objective function that minimizes pricing errors. Empirically, we find asset pricing and investment improvements using individual stocks and test portfolios for in-sample and out-of-sample analysis.