New dataset for forecasting realized volatility: Is the Tokyo stock exchange co-location dataset helpful for expansion of the heterogeneous autoregressive model in the Japanese stock market?

This study analyzes the importance of the Tokyo Stock Exchange Co-Location dataset (TSE Co-Location dataset) to forecast the realized volatility (RV) of Tokyo stock price index futures. The heterogeneous autoregressive (HAR) model is a popular linear regression model used to forecast RV. This study expands the HAR model using the TSE Co-Location dataset, stock full-board dataset and market volume dataset based on the random forest method, which is a popular machine learning algorithm and a nonlinear model. The TSE Co-Location dataset is a new dataset. This is the only information that shows the transaction status of high-frequency traders. In contrast, the stock full-board dataset shows the status of buying and selling dominance. The market volume dataset is used as a proxy for liquidity and is recognized as important information in finance. To the best of our knowledge, this study is the first to use the TSE co-location dataset. The experimental results show that our model yields a higher forecast out-of-sample accuracy of RV than the HAR model. Moreover, we find that the TSE Co-Location dataset has become more important in recent years, along with the increasing importance of high-frequency trading.