Scholarlink

Customer churn in the banking sector represents a critical challenge due to intense market competition and the high costs associated with customer loss. This article presents a multi-criteria comparison of four advanced supervised models: Neural Network, XGBoost, LightGBM, and CatBoost, which were applied to the churn prediction problem using real banking datasets. The evaluation focuses on three fundamental dimensions: predictive performance, computational efficiency, and model explainability. In terms of predictive performance, standard metrics such as Accuracy, Precision, Recall, and F1-Score were used, along with ROC and Precision-Recall curves. The results showed that CatBoost achieved the highest overall performance with an Accuracy of 0.868 and an F1-score of 0.596, followed by LightGBM and XGBoost, slightly outperforming the neural network in recall and overall balance. Regarding efficiency, LightGBM stood out for its fast training and prediction speed, while CatBoost demonstrated strong stability without the need for manual encoding of categorical variables. For interpretability, Shapley Additive exPlanations (SHAP) techniques were applied to analyze the contribution of key variables, which is essential in regulated banking contexts. The combination of results identified CatBoost as the model with the best balance between precision, efficiency, and explainability. Therefore, the multi-criteria comparative approach provides a robust basis for selecting predictive models in the financial industry, maximizing customer retention, and optimizing strategic decision-making.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-criteria Comparison of Supervised Models for Bank Customer Churn Prediction: A Performance, Efficiency, and Explainability-Based Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Michell Cárdenas,
  • Lissette Aguilar,
  • Fausto Cevallos,
  • E. Fabián Rivera

摘要

Customer churn in the banking sector represents a critical challenge due to intense market competition and the high costs associated with customer loss. This article presents a multi-criteria comparison of four advanced supervised models: Neural Network, XGBoost, LightGBM, and CatBoost, which were applied to the churn prediction problem using real banking datasets. The evaluation focuses on three fundamental dimensions: predictive performance, computational efficiency, and model explainability. In terms of predictive performance, standard metrics such as Accuracy, Precision, Recall, and F1-Score were used, along with ROC and Precision-Recall curves. The results showed that CatBoost achieved the highest overall performance with an Accuracy of 0.868 and an F1-score of 0.596, followed by LightGBM and XGBoost, slightly outperforming the neural network in recall and overall balance. Regarding efficiency, LightGBM stood out for its fast training and prediction speed, while CatBoost demonstrated strong stability without the need for manual encoding of categorical variables. For interpretability, Shapley Additive exPlanations (SHAP) techniques were applied to analyze the contribution of key variables, which is essential in regulated banking contexts. The combination of results identified CatBoost as the model with the best balance between precision, efficiency, and explainability. Therefore, the multi-criteria comparative approach provides a robust basis for selecting predictive models in the financial industry, maximizing customer retention, and optimizing strategic decision-making.