Optimizing credit risk prediction in the financial sector using boosting algorithms: a comparative study with financial datasets

Abstract

This article analyzes credit risk in the financial sector and proposes a methodology to improve its prediction accuracy using boosting algorithms such as XGBoost, LightGBM, and Boosted Random Forest. Datasets from the UCI Machine Learning Repository were used, including Statlog German Credit Data, Australian Credit Approval, and Bank Marketing. The methodology involved feature engineering, exploratory data analysis, and hyperparameter tuning. Additionally, a complementary strategy using K-means clustering was implemented to enhance the data. The results show that XGBoost outperforms the other models in various scenarios, and boosting-based methods deliver better performance than traditional approaches like decision trees and factorization machines—offering valuable insights for financial institutions.

Este artículo analiza el riesgo crediticio en el sector financiero y propone una metodología para predecirlo con mayor precisión mediante algoritmos de boosting como XGBoost, LightGBM y Boosted Random Forest. Se utilizaron datasets del repositorio UCI como Statlog German Credit Data, Australian Credit Approval, Bank Marketing, entre otros, aplicando técnicas de feature engineering, análisis exploratorio y ajuste de hiperparámetros. Además, se incorporó una estrategia adicional con K-means para enriquecer los datos. Los resultados muestran que XGBoost supera a los demás modelos en distintos escenarios, y que los métodos de boosting ofrecen mejor desempeño que enfoques tradicionales como árboles de decisión y máquinas de factorización, lo cual resulta valioso para las entidades financieras.