An explainable SHAP–LIME integrated stacked ensemble framework for bitcoin closing price forecasting

Abstract

Accurate Bitcoin price forecasting is challenging because of the highly dynamic, nonlinear, and volatile nature of cryptocurrency markets. This study proposes a comprehensive stacked ensemble framework integrating SHapley Additive exPlanations (SHAP)-based feature selection, deterministic lag-variable construction, Optuna-assisted hyperparameter optimization, and 20-fold expanding-window time-series cross-validation for Bitcoin closing-price forecasting. Multiple stacking architectures based on AdaBoost (AD), CatBoost (CB), XGBoost (XG), Random Forest (RF), Inverted Transformer (I-TRAN), and alternative meta-learners were systematically evaluated against Autoregressive Integrated Moving Average (ARIMA), Seasonal Autoregressive Integrated Moving Average (SARIMA), Historical Mean Model (HMM), and Random Walk Method (RWM) baselines. The AD + XG + RF + CB(M) architecture achieved the best performance, with Mean Absolute Error (MAE) of 169.65, Root Mean Square Error (RMSE) of 241.93, Weighted Mean Absolute Percentage Error (WMAPE) of 0.35, and Coefficient of Determination (R2) of 0.975. It also obtained Rank 1 across all evaluation metrics. Seed-sensitivity analysis confirmed model stability, while Local Interpretable Model-agnostic Explanations (LIME) demonstrated the dominant contribution of recent lagged target and price variables. The findings establish the proposed ensemble as an accurate, robust, and interpretable framework for short-term Bitcoin forecasting.