The Superiority of Direct Neuro Volatility Forecasts Over GARCH and Machine Learning Forecasts for Financial Assets

Traditional volatility forecasting often relies on generalized autoregressive conditional heteroskedasticity (GARCH) models. Moreover, conditional variance forecasts are first obtained, and then the volatility forecasts are obtained as the square root of conditional variance forecasts. However, these methods assume that log returns are either normally distributed or <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$t$</tex> distributed and are less efficient for heavy-tailed <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$t$</tex> distributions with excess kurtosis having infinite asymptotic variance. Unlike the existing work, the novelty of this paper is to use recently proposed data-driven neuro volatility models to obtain direct volatility forecasts utilizing neural network models (NNAR, RNN, LSTM, and GRU). The proposed approach eliminates the need to model conditional variance, thereby addressing inefficiencies in traditional methods. The forecast accuracies of single-layer, twolayer, and bidirectional neural networks are compared against GARCH-based models using the CBOE Volatility Index (VIX) and Bitcoin (BTC). The results demonstrate that the data-driven neuro volatility models significantly outperform the GARCH models.