Out of the Box: Deep Learning for Volatility Forecasting

We inspect how effective out of the box multivariate time series Transformer models are in forecasting a volatility index and how viable they are for future general use for volatility index forecasting by non-machine learning engineers. We analyze the performance of a traditional ARIMA-GARCH econometric model and compare it with the performance of an out of the box Light Gradient Boosting Machine and an out of the box Multivariate Time Series Transformer machine learning model. We implement our machine learning models with no hyperparameter tuning and little feature generation. Despite the low level of model refinement, the multivariate time series transformer out performed the GARCH model and demonstrated a greater ability to predict when there would be changes in volatility. This is likely due to the use of positional encoding that allows the multivariate time series transformer to learn the relationship between nearby datapoints. We propose the development of more accessible time series transformers for use in volatility forecasting.