Enhanced BSE SENSEX Forecasting Through Technical Indicator Integration and PSO-Based Feature Selection

Forecasting a stock index is difficult because daily prices are noisy, trending and sensitive to external events. Different research is performed by using different Machine Learning (ML) techniques. The prediction is more complex due to un-consistent nature of stock market. This need a enhance method that use the ML technique with data enhancement for robust0 prediction. This study develops a reproducible, leakage-aware machine-learning pipeline that predicts the next-day closing value of the BSE SENSEX from the current day's open, high, low and close prices, using 1,113 daily observations from 1 April 2022 to 8 July 2026. Prices were Min-Max normalized, the target was the closing price shifted forward by one day, and the data were split chronologically (earliest 80% for training, latest 20% for testing) so that no future information reaches the model. A Random Forest Regressor (100 trees, maximum depth 10) obtained a mean absolute error (MAE) of 553.39 points, a root mean squared error (RMSE) of 725.71 points and an R² of 0.9625 on the test set, an average error of about 0.72% of the latest close, and forecast a next-day close of 77,428.05 against a latest close of 76,922.64. Adding four technical indicators (SMA, EMA, RSI and MACD) did not improve accuracy (MAE 563.68, R² 0.9605), and Particle Swarm Optimization (PSO) tuning (70 trees, depth 5) only restored baseline-level performance (MAE 549.51, R² 0.9625). The predicted rise of 505.41 points is smaller than the model's average error, the high R² largely reflects day-to-day price persistence, and tree models cannot extrapolate beyond their training range. A simple Random Forest on daily prices therefore tracks the index level well, but neither indicators nor tuning adds measurable value.