Gwo-Lstmx : Grey Wolf Optimized Lstm with Exogenous Factors for Bitcoin Price Forecasting

This study compares the LSTM neural network and ARIMAX GARCHX models for forecasting the multivariate time series of Bitcoin prices. Predictions for Bitcoin prices were made using long-term observed data from 01/07/2019 to 31/12/2022. The prediction models utilized social-based data, including Twitter sentiment index, Twitter volume, and Google trends search index, as well as economic data such as stock indexes, Federal funds rate, gold price, and oil price. To ensure the stationarity of the data, we performed data analysis and applied log transformation, scaling, and differencing techniques. The results of the study revealed the following: 1) The inclusion of exogenous factors in the ARIMA and GARCH models yielded superior prediction results compared to the univariate model, supporting previous limited studies. 2) The inclusion of exogenous factors in the ARIMA and GARCH models yielded better prediction results compared to the univariate model, confirming findings from previous limited studies. 3) The GWO-LSTMX hyperparameters, obtained through tuning the GWO (Grey Wolf Optimization) algorithm, enhanced the performance of the LSTMX model and reduced overfitting. 4) Although the "black-box" nature of machine learning made it challenging to comprehend the implications of exogenous factors in the model, permutation importance could serve as an alternative solution. In our case, the authors discovered that all factors significantly contributed to the model, with stock indexes exerting the most substantial influence on Bitcoin prices