Comparing Random Forest and SHAP Feature Selection for LSTM-Based Stock Price Prediction

Feature selection is often used in stock price prediction, but its results remain mixed. This study analyzes the effect of feature selection based on Random Forest native feature importance (RF Importance) and SHAP on LSTM performance in predicting ADRO, ASII, and TLKM stock prices. Daily data from May 2021 to May 2026 were processed into 20 relative features from price, volume, technical indicators, and the Jakarta Composite Index (JCI), with return as the target. Both methods used the same Random Forest model and a mean-score threshold, while the LSTM was evaluated with five seeds and compared with a naive model. Feature selection reduced the input from 20 to 9 - 12 features (40% to 55%) and the LSTM parameters from 125,761 to 120,129 - 121,665 (3.3% to 4.5%). The MAPE of all schemes ranged from 1.76% to 1.90%, close to the naive model (1.72% to 1.87%), with differences between schemes of no more than 0.07 percentage points. Only for ADRO did the all-features scheme perform clearly better. Both methods retained the volume and JCI features and removed redundant trend indicators, with SHAP selecting fewer features. Therefore, the main benefit of feature selection was a simpler model input, not better prediction accuracy.