Predicting High-Frequency Industry Returns: Machine Learners Meet News Watchers
43 Pages Posted: 8 Oct 2020
Date Written: September 26, 2020
This paper uses machine learning-based as well as fundamental-driven, news-based approaches to uncover patterns of high-frequency return predictability for sector exchange-traded funds (ETFs). A LASSO predictor that aggregates high-frequency price movements of a broad universe of individual stocks predicts ETF returns out-of-sample. The news-driven return on ETF constituent firms positively predicts ETF returns, but the component of ETF returns orthogonal to the news return negatively predicts them. These different signals contain independent information, and have different strengths, with the LASSO predictor providing continuous flows of information most powerful during trading hours and the news return offering sporadic information particularly useful during market close. A composite signal combining all three signals with Gradient Boosted Regression Trees (GBRT) has very strong power to forecast ETF returns, especially during the COVID-19 pandemic.
Keywords: Industry Return, ETF, Return Predictability, LASSO, Boosted trees, News
JEL Classification: G10, G14, G40
Suggested Citation: Suggested Citation