European Volatility Under the Microscope: Applied Machine Learning for Volatility Forecasting

Sammanfattning

This paper examines the out-of-sample forecasting performance of twenty model specifications, ranging from HAR-type benchmarks, regularized linear models, and tree-based ensembles to feed-forward neural networks, in predicting one-day-ahead realized volatility on the EURO STOXX 50 index. Using high-frequency intraday trading data spanning the period from January 1999 to August 2020, we compare model performance across a parsimonious set of lagged realized variance components and an extended set augmented with macroeconomic and financial predictors motivated by our European setting. Forecasts are evaluated under both MSE and QLIKE loss functions. We find that the macro-financial predictors enable machine learning models to exploit additional signals, but the gains differ markedly across model classes and loss functions. Tree-based ensembles improve forecast accuracy under both losses, whereas regularized linear models and neural networks deliver gains under MSE only. Two specifications emerge as robust across feature sets and loss functions; HARQ on the parsimonious feature set, and Random Forest on the extended feature set. We identify the lagged realized variance components and the VSTOXX index as the main drivers for predicting next-day volatility across all four model categories.