A Strategic Management Accounting Framework for Employee Attrition Prediction using Advanced Machine Learning and XAI Techniques

The high attrition of employees causes a huge disruption in the productivity of the organization and increases the cost of hiring new talent. Therefore, accurate prediction of attrition proactively is essential for sustainable management of the enterprise. Traditional machine learning approaches are often challenged to achieve high prediction accuracy due to extreme imbalance of classes in datasets, lack of proper feature engineering and poor hyperparameter tuning across base classifiers. Here, we propose an optimized prediction framework on the IBM HR analytics dataset. Following advanced feature engineering and target balancing using SMOTE, we benchmarked eleven base classification algorithms. Crucially, we hyper-tuned a tree-based configuration, extending CatBoost to 500 iterations (depth=7, l2_leaf_reg=5) and expanding Extra Trees to full leaf node purity with entropy selection. Then, a dynamic soft-voting ensemble architecture was implemented to automatically aggregate the predictions from the top performing base models based on cross validated accuracy ranks. The experimental results show that CatBoost and Extra Trees are the best individual models, with accuracies of 84.16% and 83.26%, respectively. The proposed dual-model setup, Ensemble a+b (CatBoost + Extra Trees), outperformed all of the architectures with a maximum classification accuracy of 85.07%, an optimum precision of 72.37%, and a minimum Mean Square Error (MSE) of 0.1493. This study shows that tree ensembles optimized by a strategy and dynamic soft voting significantly reduce errors of prediction, giving organizations a reliable framework to intervene early in attrition.