Dynamic bagging-based ensemble method for machine-learning model selection in bitcoin price prediction

This study proposes a dynamic bagging ensemble that re-weights six baseline forecasters—ARIMA, SARIMA, ARIMAX, ANN, LSTM and BiLSTM—each day using rolling accuracy (MAE, MAPE, RMSE, R²), directional-accuracy and F1 scores derived from the most recent lookback window, while retaining static inverse-error weights learned during walk-forward cross-validation. When applied to out-of-sample BTC-USD data from six major exchanges (December 1, 2024 – April 1, 2025), the system achieves an average inference-stage MAE = 1545 USD, MAPE = 1.64%, RMSE = 2 060 USD, R² = 0.912, DA = 0.930 and F1 = 0.920, improving MAE by roughly 15% and DA by 3 - 5 percentage point relative to the best single-model benchmark. The ensemble also delivers a positive daily Sharpe ratio of ≈ 0.11 - 0.13 versus the 3-month U.S. Treasury-bill rate, whereas most individual learners turn negative under heavy-tuning scenarios. Selection-frequency analysis further reveals that statistical models dominate at short lookbacks (7–15 days) while LSTM-based networks gain prominence at 30-day horizons, underscoring the framework’s ability to adaptively allocate forecast weight to the most contextually effective model and thus provide a robust, low-latency decision-support tool for real-time Bitcoin trading.