HyBiLSTM: Multivariate Bitcoin Price Forecasting Using Hybrid Time-Series Models With Bidirectional LSTM

Abstract

Despite their popularity in recent studies, most hybrid models that exploit the advantages of both classical time series and deep learning models were conducted in univariate forecasting context. For econometric domain which exogenous factors play a crucial role, more studies in multivariate forecasting is essential and should be encouraged. Thus, contributing to hybrid multivariate forecasting literature, a hybrid model named HyBiLSTM was proposed. The algorithm began with ARIMAX GARCHX model forecasting, followed by second forecasting of model residual using Grey Wolf Optimizer based hyperparameters Bidirectional LSTM model. With residuals instead of original multivariate features, LSTM can avoid to processes each feature independently and therefore, reducing convergence complexity and execution time. The final forecasting results was compounded from both models. Three quantitative measurements, Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE), were used to evaluate the established models using the historical daily runoff social and economic based data (01/07/2019-31/12/2022). The findings showed 1) the addition of exogenous factors improved the performance of ARIMA and GARCH models; 2) BiLSTM outperformed other LSTM variants when was integrated with ARIMAX GARCHX model; 3) Using SHAP, Bitcoin price was influenced by stock price, Twitter volume, gold price, and Twitter sentiment index; and 4) structural break had significant effect on forecasting. Other than expanding the literatures regarding hybrid models in multivariate context, this study provides practical contribution for investors by analyzing the factors that the investors can use as an early warning for Bitcoin price fluctuation.