This study thoroughly examines the impact of feature selection methods used for stock market index prediction on model performance using XGBoost and Boruta approaches. Using nine feature sets created from different combinations of these two feature selection methods, Long Short-Term Memory (LSTM) and LSTM with Batch Normalization (LSTM-BN) models were trained to predict the closing values of the Standard & Poor’s 500 (S&P 500), Shanghai Stock Exchange Composite Index (SSEC), and Shenzhen Stock Exchange Component Index (SZSE). The analysis was carried out by evaluating the models using metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared to ensure a comprehensive understanding of model performance. The results indicate that feature selection has a decisive effect on the performance of the prediction models. Specifically, it was observed that LSTM models trained with the XGBoost-Top30 feature set generally achieved the lowest error rates (approximately RMSE: 0.02, MAE: 0.01, \(R^2\) : 0.99), suggesting that this particular feature set effectively captures the most relevant features needed for accurate predictions. This highlights the potential of the XGBoost-Top30 feature set to enhance model accuracy and reliability significantly. These findings underscore the critical role of feature selection in stock market index prediction. By effectively identifying and utilizing the most pertinent features, the accuracy of predictive models can be substantially improved. Furthermore, the study highlights the importance of investigating more comprehensive feature selection techniques in future studies. Developing and refining these techniques will be crucial for advancing predictive modeling in the financial domain.