Title: Deep learning alpha
Authors: Guanhao Feng - City University of Hong Kong (Hong Kong) [presenting]
Nicholas Polson - University of Chicago (United States)
Jianeng Xu - University of Chicago (United States)
Abstract: The goal is to push the classical long-short portfolio asset pricing framework to the future of artificial intelligence. Sorting securities on firm characteristics and constructing long-short portfolios is a tradition to both the asset pricing academia and hedge fund industry. Sorting is a nonlinear activation that can be built within a deep learning architecture and works for the unbalanced panel data with various missing values. In a general setup, we develop a multi-layer neural network to augment additional long-short latent factors to a factor model like CAPM. The approach explores the firm characteristic search space via various nonlinear transformations with an economic objective: to eliminate mispricing alphas. Our algorithm provides a joint estimation for the augmented linear factor model and the underlying neural network. To illustrate the method, we design our long-short latent factor construction in a train-validation- test framework. From an empirical perspective, we perform an out-of-sample study to analyze Fama-French factors in both the cross-section and time series. The finding is a significant forecasting improvement by adding nonlinear signals from firm characteristics.