Machine learning for realised volatility forecasting

We assess the predictive power of machine learning (ML) models for forecasting realised volatility using information from HAR model variables, limit order book (LOB) data, and news sentiment. Training and robustness checks on nearly seven million ML models show that high-dimensional ML models outperform HAR models in 90% of the out-of-sample period, except during extreme volatility. Explainable AI analysis identifies mid prices, mean bids, and mean asks as key predictors. Notably, incorporating ML into ensemble frameworks enhances HAR model performance, though caution is needed when using ML models as direct substitutes, since they may yield unreliable forecasts under certain market conditions.