Neural Networks in Forecasting Financial Volatility

In 2020s, the state of the art (SOTA) in financial volatility forecasting is underpinned by deep learning (DL). Despite this, forecasting methods in practice tend to be dominated by their more traditional counterparts (e.g., Generalised Auto-Regressive Conditional Heteroscedasticity (GARCH) models) or relatively simple neural networks (NN), leaving much of DL unexplored. Hence, this study experimented the power of DL in forecasting financial volatility and expedited further progress in such multidisciplinary DL applications to quantitative finance by releasing open-source software and proposing a shared task. We compared the financial forecasting ability of the SOTA methods used to more recent DL work, proceeding from simpler or shallower to deeper and more complex models. Specifically, the volatility of five assets (i.e., S&P500, NASDAQ100, gold, silver, and oil) was forecast with the GARCH models, multi-layer perceptrons, recurrent NNs, temporal convolutional networks, and Temporal Fusion Transformer. The results indicated that in almost all cases, DL models forecast volatility with less error than the SOTA models in financial volatility research. These experiments were repeated and the difference between competing models was shown to be statistically significant, therefore encouraging their use in practice.