Short- to Long-Term Realized Volatility Forecasting using Extreme Gradient Boosting

Abstract

We adopt Extreme Gradient Boosting (XGBoost) to forecast realized volatility. This is motivated by XGBoost's strong forecasting performance in other forecast applications and its ability to capture non-linearities, a feature that is also frequently reported in the context of realized volatility. We examine the forecasting precision of linear and non-linear XGBoost models for different forecast horizons and compare it to that of Long Short-Term Memory (LSTM) networks as well as heterogeneous autoregressive (HAR) models. We find that XGBoost exhibits a better forecast performance. In particular, XGBoost models significantly outperform both HAR and LSTM models for one-step ahead predictions. For longer forecast horizons, linear models such as XGBoost with linear base learners perform better than non-linear specifications, suggesting that accounting for non-linearities is only important if short-term forecasts are of interest.

Keywords: Volatility Forecasting, Realized Variance, High-frequency Data, Extreme Gradient Boosting, LSTM, Machine Learning

JEL Classification: C58,C55,C50