A Hybrid Model Integrating LSTM and Garch for Bitcoin Price Prediction

Abstract

Due to the nonlinearity and highly volatile dynamics of the price data of cryptocurrency, classic parametric models show limited success in tracking and prediction. With the rise of deep learning recently, various researches on forecasting the price of cryptocurrency using deep neural network have reported encouraging results in the cases of low volatility. In this study, we propose a hybrid approach which combines the advantages of non-stationary parametric models such as Generalized Autoregressive Conditional Heteroskedasticity (GARCH) with the nonlinear modelling potential of Long-Short Term Memory (LSTM) neural networks. The results show that our hybrid model has a similar predictive performance in terms of MSE, MAE and RMSE but higher metric scores in precision, accuracy and F1 score under optimal hyperparameters. This study reveals that the combination of parametric models like GARCH with deep neural network may come up with better results in cryptocurrency price forecasting especially in the case of highly volatile data or when short data sequences are available. Moreover, the proposed framework can be used also in other applications where high volatility and scarcity of data are the main characteristics.