A hybrid deep learning model for Bitcoin price prediction: data decomposition and feature selection

Bitcoin has received a great deal of attention as a highly volatile asset with investors attempting to profit from its dramatic price fluctuations. We develop a hybrid deep learning model based on feature selection in different frequency domains to enrich the literature of Bitcoin price prediction. Indicators such as Technology, Economy, Green Finance and Media Attention are considered. We first decompose all the data into different frequencies through CEEMDAN approach, and then the data at the same frequency are integrated into a Random Forest model to reduce the subset of potential predictors by measuring the importance of different factors. Finally, the selected factors are put into the LSTM/GRU to make the prediction of different components of Bitcoin prices at the same frequency, and aggregate together to obtain the predicted Bitcoin prices. The empirical results show that our proposed model outperforms the benchmark models, which is verified by MCS test. The proposed hybrid method obtains much higher return on investment in simulated trading than other benchmark models. Our study inspired the investors to accurately predict Bitcoin price and dig possible relationships between different assets and its determinants in frequency domain.