Comparative Study of Hybrid ARIMA-LSTM and ARIMAX-LSTM for Bitcoin Forecasting with Data Partitioning

The extreme volatility of Bitcoin prices poses significant challenges for accurate forecasting using conventional models. While ARIMA excels at capturing linear trends, it struggles with non-linear dynamics; conversely, LSTM networks can model non-linearity but often overfit noisy data. To address these limitations, this study investigates six forecasting configurations: standalone ARIMAX, standalone LSTM, and four hybrid ARIMA/ARIMAX-LSTM models employing both single-split and two-stage split strategies. A comprehensive out-of-sample evaluation on daily Bitcoin closing prices reveals that the two-stage split hybrid ARIMA-LSTM achieves a remarkable MAPE of 2.60%, outperforming all other configurations. The results demonstrate that residual structure and strategic data partitioning critically influence hybrid model performance by enhancing residual learnability. These findings offer practical guidance for researchers and practitioners designing robust forecasting pipelines for highly volatile financial markets.