Russian Stock Market Realized Volatility Forecasting Using Machine Learning

Volatility of stock returns is one of the most important and commonly used measures of the asset risk, and volatility forecasting is proved a challenging task that has been widely researched. Realized volatility approach has become a preferred one among practitioners and researchers because it directly evaluates volatility from high-frequent market data and tends to provide precise forecasts of the future volatility. Since the baseline heterogeneous autoregressive model of realized volatility was introduced, substantial research has been devoted to improving its explanatory and predictive capabilities. Massive studies use a vast variety of additional variables to explain the changes in volatility on the markets, including financial indicators of firms or markets, various spillover effects, calendar effects, or macroeconomic indices. Another potentially, yet not so well studied at the moment, approach to improve the predictions is usage of nonlinear models, particularly machine learning algorithms. We combine the two approaches and study the predictive performance of selected machine learning algorithms relative to the benchmark model in the task of prediction of realized volatility of stock returns on the Russian stock market based on both historical values of the volatility itself, and a broad set of potentially informative financial and macroeconomic indicators. We find that both the traditional methodology and the machine learning approach provide us with reasonable predictive power in terms of RMSE of realized volatility forecasts on rolling training samples, with the machine learning algorithms that utilize additional exogenous regressors generally outperforming the traditional baseline autoregressive model. Particularly, Lasso regression is shown to be rather efficient both in terms of relatively easy implementation, and precision of forecasts. More complicated algorithms (Random Forest, Gradient Boosting, Long Short-Term Memory) are very promising, yet we show that to benefit from them, it requires to tune them very finely and re-train rather often, which is a computationally demanding task.