High-precision forecasting of Indian stock market indices using weighted ensemble of hyperparameter-tuned LightGBM models with diverse technical indicators

Forecasting financial markets is inherently challenging due to their high volatility, non-stationary behaviour, and the complex interplay of multiple influencing factors. Exact prediction requires models that can adapt to dynamic patterns while maintaining stability and generalization. This study proposes a robust financial market forecasting (FMF) framework that leverages the predictive power of Light Gradient Boosting Machine (LightGBM) enhanced through multiple hyperparameter optimization strategies. The model is applied to major Indian stock indices from the National Stock Exchange (NSE), including NIFTY50, NIFTYMIDCAP50, and NIFTYBANK. Four distinct optimization algorithms such as Bayesian Optimization (BO), Ant Colony Optimization (ACO), Random Search, and Teaching – Learning-Based Optimization (TLBO), are systematically employed to identify optimal configurations. Additionally, a weighted average ensemble strategy integrates the individually optimized models to reduce variance and improve generalization. The proposed framework is evaluated using standard regression metrics such as MSE, RMSE, MAE, MAPE, R², and NRMSE, and compared against baseline machine learning models. The proposed ensemble LightGBM achieved the best performance on NIFTY50 data, with RMSE = 105.192, MAE = 85.067, MAPE = 0.352, R² = 0.942, and NRMSE = 0.004, outperforming all baseline models. These results confirm that combining multi-strategy hyperparameter tuning with ensemble learning enables superior accuracy and robustness in FMF.