XGBoost-LSTM for Feature Selection and Predictions for the S&P 500 Financial Sector

Abstract

Financial industry researchers have long been committed to identifying factors that can predict trends in the financial sector of S&P 500, despite these factors often being difficult to discover. This article, through the combination of the Xgboost regressor and the shap summary plot, has mined and continuously optimized a potential excellent factor combination. Also, by utilizing the Xgboost regressor and LSTM models, it has achieved good prediction accuracy on the test set. This research gets the following results: First, the Xgboost regressor, in combination with Shap, has identified the seven most excellent factors from an initial combination of nine factors. Second, after imparting the final seven features to LSTM, the MSEs of the predictions made by Xgboost regressor and LSTM are 0.0003 and 0.0004, while the running times for Xgboost regressor and LSTM are 27 minutes and 16 minutes. Consequently, these results indicate that in the future predictions of finance sector index, investors may use the Xgboost-LSTM model for selecting effective factors and making accurate predictions efficiently.