Hybrid deep learning and GARCH-family models for forecasting volatility of cryptocurrencies

Abstract

The combination of Deep Learning and GARCH-type models has been proved to be superior to the single models in forecasting of volatility in various markets such as energy, main metals, and especially stock markets. To verify this hypothesis for cryptocurrencies market, we constructed various Deep Learning models based on Feed Forward Neural Networks (DFFNNs) and Long Short-Term Memory (LSTM) networks and evaluated their performance in forecasting the volatility of 27 cryptocurrencies. Then, different hybrid models were built in which the outputs of three GARCH-type models, namely GARCH, EGARCH, and APGARCH, with three different assumptions for the residuals’ distribution were fed into the DFFNN and LSTM networks. In other words, GARCH-type models were utilized as feature extractors and the deep learning models leveraged a sequence of extracted features as their inputs to produce the volatility of the next day. Our findings revealed that not only the deep learning models improve the forecasts of GARCH-type models with any distribution assumption, the forecasts of GARCH-type models as informative features can significantly increase the predictive power of the studied deep learning models; namely, the DFFNN and LSTM models.