A Robust Framework for Stock Price Prediction via BorutaSHAP-Based Feature Selection and Ensemble Deep Learning

Abstract

Stock price prediction remains challenging due to market noise, non-stationarity, and complex temporal dependencies. This study proposes a robust hybrid framework to enhance predictive performance, stability, and interpretability in financial time series forecasting. The framework integrates BorutaSHAP-based feature selection with variance inflation factor (VIF) filtering to construct a relevant and non-redundant feature set. To better capture temporal dynamics, time-series decomposition is combined with Moving Block Bootstrap (MBB) resampling, preserving local dependencies while increasing data diversity. The predictive stage employs a bagging-based ensemble of deep learning models, including LSTM, GRU, Conv-LSTM, and RNN architectures, with XGBoost as a benchmark. Hyperparameters are optimized using Bayesian Optimization (BO). In addition to predictive accuracy, the framework incorporates SHAP-based analysis to investigate the impact of hyperparameters on model performance and behavior. The proposed approach is evaluated on financial time series from the S&P 500 index. Experimental results demonstrate improved robustness and generalization under noisy market conditions, with deep learning-based ensembles achieving a competitive balance between accuracy and stability. Overall, the findings indicate that integrating feature selection, resampling, and ensemble deep learning can enhance the reliability of stock price prediction.