A Study of Machine Learning's Applicability in Realized Volatility Modelling

Abstract

Modelling realized volatility (RV) has been a prominent topic in the financial literature for decades. As of most recently the financial industry has started shifting more towards machine learning algorithms to model RV, moving away from traditional time series models. This study explores the modelling of one-day-ahead RV in three major stock markets: the US, UK and Japan. Traditional models, including the Heterogeneous Autoregressive (HAR) model, HAR with exogenous variables (HAR-X), Random Forest, and shrinkage models, along with innovative machine learning approaches like multitask learning and online learning, are employed for each market. For the Japanese and UK markets, all modelling approaches were found to significantly outperform the HAR model, while for the US market the HAR model was found to outperform most of the machine learning models. Additionally, the multitask learning model was found to perform just as well as the existing modelling approaches, suggesting no particular benefits from training the models simultaneously. Encouragingly, the online learning model is found to be the best performing model across all markets. In particular, the findings are emphasized by the model’s adaptability to changing environments, an aspect which is beneficial due to the dynamic nature of financial markets.