Momentum in Machine Learning: Evidence from the Taiwan Stock Market

Abstract

We revisit 86 asset pricing anomalies in the Taiwan stock market and find that long-short portfolio strategies based on machine-learning methods bring substantial benefits. For example, neural networks and partial least squares generate long-short returns ranging from 1.20% to 1.50% per month. More importantly, five of the top 20 influential return predictors are momentum-related variables. This result provides novel evidence to the momentum literature given that the Taiwan stock market is one of the few exceptions to the momentum anomaly. In contrast with this conventional view, we show that momentum contributes to stock return predictability when adopting machine-learning models.

Keywords: Momentum, Asset pricing anomalies, Stock return predictability, Machine learning, Variable importance

JEL Classification: G11, G14, G32, G35, G40