A Hybrid Ensemble Method with Focal Loss for Improved Forecasting Accuracy on Imbalanced Datasets

The inherent complexity and dynamic characteristics of diverse datasets present significant challenges for achieving high predictive accuracy in forecasting tasks. This study tackles these challenges by implementing a hybrid ensemble model aimed at enhancing predictive performance across imbalanced datasets. Using data from a competitive data source, the approach integrates LightGBM, XGBoost, and Logistic Regression models within a weighted ensemble framework to improve overall prediction accuracy. Data preprocessing techniques, including KNN imputation, Z-score normalization, and SMOTE, are employed to handle missing values, outliers, and class imbalances, ensuring a robust input for model training. The ensemble framework incorporates a Focal Loss function to specifically address class imbalances and refine prediction precision. Comparative analyses reveal that the proposed ensemble model consistently outperforms individual models in terms of accuracy, precision, recall, and AUC. This study offers a versatile and reliable solution for forecasting challenges, demonstrating enhanced robustness and broad applicability across domains.