More is better? The impact of predictor choice on the INE oil futures volatility forecasting

This paper aims to address the predictor choice issue in forecasting volatility of INE oil futures by a comprehensive comparative study with a large number of predictive variables and applying machine learning models along with their interpretability tools. The main finding is that the selection of predictors is crucial for improving volatility forecasting accuracy, but it is not always the case that including more predictive variables leads to better forecasting results, even for machine learning models. Specifically, this paper has five major findings: (1) A few variables can significantly improve forecasting accuracy independently, but their contribution is limited. (2) Increasing the number of predictors from specific categories (market sentiment indicators, crude oil futures prices from other exchanges, and energy market indicators) helps to enhance forecasting accuracy. (3) Low-frequency variables have a weak effect on improving the daily volatility. (4) Ensemble tree models perform better than traditional machine learning models based on variable selection with dynamic parameter optimization, even without much parameter tuning. The above findings still hold true under a series of robustness tests and economic value assessments. These findings provide substantial evidence for addressing the issues of model and variable choice in crude oil futures volatility forecasting. •Individual predictors' contribution to improving volatility forecasting performance of the INE oil futures is limited to 5%.•Adding more predictors boosts accuracy,but the significant improvement comes from predictors of specific categories.•In HAR-RV and machine learning models, low-frequency variables have a weak impact on improving RV forecasting performance.•The best forecasting model doesn't need many variables, and key categories of predictors ensure optimal performance.