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

Putri Auliana Rifqi Mukhlashin (Department of Statistics, Faculty of Mathematics and Science, IPB University, West Java)
Anwar Fitrianto (Department of Statistics, Faculty of Mathematics and Science, IPB University, West Java)
Agus M Soleh (Department of Statistics, Faculty of Mathematics and Science, IPB University, West Java)
Wan Zuki Azman Wan Muhamad (Institute of Engineering Mathematics, Universiti Malaysia Perlis, Arau)

Journal of Accounting and Investment
Vol 24, No 2: May 2023

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