Diverse Minds Think Alike: A SHAP-Guided RNN for Stock Price Forecasting with Smart SMA Optimization

Stock market forecasting is challenging due to volatility, complex temporal dependencies and nonlinear patterns. Many models underutilize features, lack interpretability and fail to leverage model diversity in ensembles. This paper proposes an explainable ensemble framework combining attention enhanced recurrent networks (LSTM, GRU, BiLSTM) with engineered technical and sentiment features. Optimal Simple Moving Average (SMA) windows are selected via a hybrid A* search and dynamic programming method. SHapley Additive exPlanations (SHAP) quantify feature importance and inter model diversity, guiding a dynamic weighting scheme for ensemble aggregation. These weighted outputs along with base model predictions are input to an XGBoost meta-learner for final forecasting. Experiments on real world stock and sentiment data show the method performs better than individual models and static ensembles achieving an MSE of 1.0445, MAE of 0.6397 and $R^{2}$ of 0.9861. The framework advances explainability, temporal modeling and diversity aware ensembling for more reliable stock price prediction.