Deep Neural Networks in Predicting Financial Volatility: An Application on Cryptocurrencies

The main aim of this research is to obtain daily volatility predictions of cryptocurrencies by using the Deep Belief Network – Deep Neural Network (DBN-DNN) and Backpropagation Neural Network (BPNN) algorithms as well as the traditional Generalized Autoregressive Conditional Heteroskedasticity (GARCH) and one of its asymmetric variants, the Exponential GARCH models. The research data consists of daily price series of the most well-known two cryptocurrencies (Bitcoin and Ethereum) for the period of 4 July 2022 and 3 February 2024. Performance of the out-of-sample volatility forecasts of the models are compared by using three different metrics, that are Root Mean Squared Error (RMSE), R squared (R2) and Mean Absolute Error (MAE). According to the obtained performance measures estimated from the out-of-sample period of 11 October 2023 and 3 February 2024, DBN-DNN and BPNN yielded the best RMSE and MAE values of Bitcoin, respectively. On the other hand, DBN-DNN obtained the best RMSE and R2 among the Ethereum volatility predictions. Overall, the findings of this research reveal a promising cryptocurrency volatility forecasting ability in favor of the applied deep neural networks, since they managed to obtain most of the best as well as the second-best performance metrics of both cryptocurrencies.