A machine learning approach to volatility forecasting

We show that machine learning (ML) algorithms improve one-day-ahead forecasts of realized variance from 29 Dow Jones Industrial Average index stocks over the sample period 2001 - 2017. We inspect several ML approaches: Regularization, tree-based algorithms, and neural networks. Off-the-shelf ML implementations beat the Heterogeneous AutoRegressive (HAR) model, even when the only predictors employed are the daily, weekly, and monthly lag of realized variance. Moreover, ML algorithms are capable of extracting substantial more information from additional predictors of volatility, including firm-specific characteristics and macroeconomic indicators, relative to an extended HAR model (HAR-X). ML automatically deciphers the often nonlinear relationship among the variables, allowing to identify key associations driving volatility. With accumulated local effect (ALE) plots we show there is a general agreement about the set of the most dominant predictors, but disagreement on their ranking. We investigate the robustness of ML when a large number of irrelevant variables, exhibiting serial correlation and conditional heteroscedasticity, are added to the information set. We document sustained forecasting improvements also in this setting.