Machine Learning for Realised Volatility Forecasting

Abstract

We examine the predictive power of machine learning (ML) models for forecasting realised volatility (RV) using different data sets tested in the literature, viz., variables from the HAR models, limit order book (LOB), and news sentiments. With 3.7 million ML models trained and robustness checked, the high dimensional ML models (viz. 147 data features for up to 21 lags) significantly outperformed HAR on 90% of the out-of-sample period when the actual volatility is not extreme. SHAP values reveal that mid prices, mean bids, and mean asks are the most useful predictive variables. The ML models also better than HAR at capturing forecast dynamics as they evolve through time.

Keywords: Realised Volatility Forecasting, Machine Learning, Big Data, Long Short-Term Memory, Heterogeneous AutoRegressive Models, Explainable AI.

JEL Classification: C22, C45, C51, C53, C55, C58