Practical Applications of The Best of Both Worlds: Forecasting US Equity Market Returns Using a Hybrid Machine Learning–Time Series Approach

Abstract

In <b><i>The Best of Both Worlds: Forecasting US Equity Market Returns Using a Hybrid Machine Learning–Time Series Approach</i></b>, published in the Spring 2021 issue of <b><i>The Journal of Financial Data Science,</i></b><b>Haifeng Wang, Harshdeep Singh Ahluwalia, Roger Aliaga-Díaz, and Joseph Davis</b> (all at<b> Vanguard</b>) revisit the long-term equity return forecasting performance of a popular time-series model based on Shiller’s cyclically adjusted price-to-earnings (CAPE) ratio. Noting that its performance had deteriorated due to structural economic changes, they considered additional variables, econometric methods, and machine learning (ML) techniques. Using their domain knowledge in the data science arena, the authors sought to improve the out-of-sample predictive accuracy of long-run stock return forecasts. They compared the performance of various approaches, including linear regressions, vector autoregressive (VAR) models, random forests, gradient boosting machines, and gated recurrent units. They found that a two-step approach that estimates a VAR model with ML techniques and uses outputs in an additive component model of returns works best. Using an ensemble of ML techniques within a robust economic framework yields more accurate long-run return predictions than using any of the methods alone.