Ensemble learning with imbalanced data handling in the early detection of capital markets

Research aims: This study aims to create an early detection model predict events in the Indonesian capital market.Design/Methodology/Approach: A quantitative comparing ensemble learning models with imbalanced data handling detected market events. used five models—Random Forest, ExtraTrees, CatBoost, XGBoost, and LightGBM—to detect by data, such as under sampling (RUS), oversampling (SMOTE, SMOTE-Broder, ADASYN), over-under (SMOTE-Tomek, SMOTE-ENN), weighted (class weight). Global regional stock markets, commodities, exchange rates, technical indicators, sectoral indices, JCI leaders, MSCI, net buys of foreign stocks, national securities, share ownership all predicted lowest return Crisis Management Protocol (CMP) binary responses.Research findings: Hyperparameters thresholds were tuned produce optimum model. The best had highest G-mean. ExtraTrees SMOTE-ENN number one-day events, a G-Mean 96.88%. LightGBM SMOTE five-day 89.21% G-Mean. With 89.49%, CatBoost SMOTE-Border was for 15-day event. In addition, SMOTE-Tomek 68.02% 30-day Further, performance evaluation scores decreased increased prediction time.Theoretical contribution/Originality: work relates more imbalance methods cases.Practitioner/Policy implication: Capital markets can indicate economic stability. Maintaining efficacy value requires system pressure.Research limitation/Implication: 1, 5, 15, 30 days ahead, assuming working days. model's forecast results are expected be utilized monitor take precautions.