Abstract Forecasting volatility is an important need in financial markets, especially for more volatile asset classes such as cryptocurrencies. Standard statistical models based on GARCH-type specifications typically will not have the capacity to specify complex nonlinearities usually present in the cryptocurrency space. This paper undertakes an in-depth evaluation of three deep learning architectures: a multilayer perceptron (MLP), long short-term memory network (LSTM), and Transformer models (including an augmentative convolution model) for forecasting realized volatility in high-frequency data. Specifically, high-frequency price data is aggregated through intra-day returns to a measure of daily realized volatility through a series of log-transformation and normalization. The architectures take a supervised rolling window framework and are evaluated using standard error metrics (MSE), heteroskedastic error measures (HAE, HSE), and risk measures (Sharpe ratio and Value-at-Risk). Results suggest that, despite the MLP and LSTM being competitive baselines, Transformer-type models, especially the convolution-augmented model, are superior in forecasting volatility spikes and modeling complex temporal structures. Plots of the forecast and actual for the three models also illustrate these points. These findings could assist in improving the way to deploy deep learning based forecasting methods into risk management and portfolio optimization contexts in computational economics.
R. Naresh (Tue,) studied this question.