Timing the factor zoo via deep learning: Evidence from China

Abstract

This paper proposes a factor timing strategy with information from 146 characteristic‐based factors and a deep learning approach to capture the nonlinear predictability. The deep learning‐based factor timing strategy generates the highest economic value compared with the unconditional and alternative linear machine learning‐based portfolios and remains robust after controlling for traditional factor models and transaction costs. With the unique market structure of the Chinese stock market, we find that mispricing‐based theory helps explain the factor timing via deep learning.