Charting new avenues in financial forecasting with TimesNet: The impact of intraperiod and interperiod variations on realized volatility prediction

Abstract

This study evaluates TimesNet model for stock realized volatility forecasting, comparing its efficacy against traditional and contemporary models across key metrics: RMSE, MAE, MAPE, and QLIKE. Despite TimesNet’s competitive performance in RMSE and QLIKE, it faces challenges in MAE and MAPE, falling behind models like HAR, NBEATSx and NHITS. Highlighting the novel application of CNN architectures beyond image recognition, this research suggests TimesNet’s potential in long-term forecasting and extreme event prediction, yet indicates the necessity for further model enhancement. The findings motivates new avenues for future research into CNN-based forecasting models in stock realized volatility forecasting by improving the current version of the TimesNet model.