Forecasting realized volatility in turbulent times using temporal fusion transformers

This paper analyzes the performance of temporal fusion transformers in forecasting realized volatilities of stocks listed in the S&P 500 in volatile periods by comparing the predictions with those of state-of-the-art machine learning methods as well as GARCH models. The models are trained on weekly and monthly data based on three different feature sets using varying training approaches including pooling methods. I find that temporal fusion transformers show very good results in predicting financial volatility and outperform long short-term memory networks and random forests when using pooling methods. The use of sectoral pooling substantially improves the predictive performance of all machine learning approaches used. The results are robust to different ways of training the models.