Time-mixing and feature-mixing modelling for realized volatility forecast: Evidence from TSMixer model

Abstract

This study evaluates the effectiveness of the TSMixer neural network model in forecasting stock realized volatility, comparing it with traditional and contemporary benchmark models. Using data from S&P 100 index stocks and three other datasets containing various financial securities, extensive analyses, including robustness tests, were conducted. Results show that TSMixer outperforms benchmark models in predicting individual stock volatility when applied to datasets with a large number of securities, leveraging its feature-mixing MLP techniques, which can properly model the financial tail dependence phenomenon. However, its superiority diminishes in datasets with fewer securities, such as stock indexes, foreign exchange rates, and commodities, where models like NBEATSx and NHITS often perform better. This indicates that TSMixer's performance is context-dependent, excelling when feature interdependencies can be fully exploited. The findings suggest that simplified neural network architectures like TSMixer can enhance forecasting accuracy in appropriate contexts but may have limitations in datasets with fewer securities.