Volatility Forecasting with Machine Learning

Volatility forecasting estimates how much an asset’s price may vary over a future horizon. It is distinct from predicting whether price will rise or fall. Machine-learning methods can use lagged realized volatility, market-wide signals, macroeconomic variables, or text sentiment. A common benchmark is the heterogeneous autoregressive realized-volatility model (HAR-RV), which uses averages of past volatility over different time horizons. Research comparing ML with linear models finds that results depend on the asset, predictors, forecast horizon, and evaluation design. One study across global equity indices found additional predictors improved some daily and weekly forecasts, but no general evidence that nonlinear ML always outperformed linear methods. A forecasting metric does not show that a strategy is profitable after costs or suitable for a particular portfolio. Researchers should use rolling or expanding out-of-sample tests, prevent look-ahead leakage, compare against simple baselines, and evaluate calibration and loss functions relevant to the decision. Market regimes change, and models can degrade during stress. Report horizons, target construction, assets, fees, and uncertainty. Do not interpret a volatility estimate as a prediction of returns or as individualized financial advice. Risk managers should specify whether the forecast informs position sizing, stress testing, or capital planning, since each task values different error costs. A model selected for one market or forecast horizon may not remain useful in another. Trading costs and liquidity can change the apparent value of a volatility signal.