Evaluación de modelo híbrido de calificación de riesgo crediticio para banca empresarial en Honduras, período de enero 2020 - octubre 2024
No Thumbnail Available
Date
2026-03-02
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Tecnológica Centroamericana UNITEC
Abstract
Esta investigación evaluó el impacto de un modelo híbrido de calificación de riesgo crediticio, que combina aprendizaje automático y reglas de negocio, aplicado a clientes de banca empresarial en Honduras entre enero de 2020-octubre de 2024. Se utilizó un enfoque cuantitativo con la metodología CRISP DM, a partir de una base de 6,903 clientes y 42 variables financieras, conductuales, demográficas y regulatorias. Los datos se dividieron en entrenamiento (2020-2022), validación (2023) y prueba out of time (2024). Se comparó un esquema de reglas, una regresión logística y modelos de ML (Random Forest y XGBoost), evaluando su desempeño mediante AUC ROC, KS y métricas operativas al umbral definido en la validación. En el conjunto de prueba, Random Forest alcanzó AUC ROC 0.87 y KS 0.58, y XGBoost AUC ROC 0.86 y KS 0.55, superando ampliamente al enfoque tradicional (AUC 0.59; KS 0.18) y elevando el recall de la clase morosa de 0.66 a 0.87, con estabilidad temporal (PSI≈0.0066) y buena calibración (Brier≈0.05- 0.06). La reducción de falsos negativos se tradujo en pérdidas evitadas estimadas en L13 millones y un ROI de ≈825.95%. Se recomendó la implementación piloto del modelo híbrido bajo un marco de gobierno de modelo acorde con Basilea III, BCBS 239 y SR 11 7.
This study evaluated the impact of a hybrid credit risk rating model that combines machine learning techniques and business rules, applied to corporate banking clients in Honduras between January 2020 and October 2024. A quantitative approach based on the CRISP DM methodology was used, drawing on a dataset of 6,903 clients and 42 financial, behavioural, demographic and regulatory variables. The data were split into training (2020-2022), validation (2023) and out of time test (2024) sets. A rule based scheme, a logistic regression and machine learning models (Random Forest and XGBoost) were compared, and their performance was assessed using AUC ROC, KS and operational metrics at the decision threshold defined in the validation stage. In the test set, Random Forest achieved an AUC ROC of 0.87 and a KS of 0.58, and XGBoost an AUC ROC of 0.86 and a KS of 0.55, clearly outperforming the traditional rule based approach (AUC 0.59; KS 0.18) and increasing the recall of the default class from 0.66 to 0.87, with temporal stability (PSI≈0.0066) and good calibration (Brier≈0.05-0.06). The reduction in false negatives translated into avoided losses estimated at L 13 million and a ROI of approximately 825.95 %. A pilot implementation of the hybrid model is recommended under a model risk governance framework aligned with Basel III, BCBS 239 and SR 11 7.
This study evaluated the impact of a hybrid credit risk rating model that combines machine learning techniques and business rules, applied to corporate banking clients in Honduras between January 2020 and October 2024. A quantitative approach based on the CRISP DM methodology was used, drawing on a dataset of 6,903 clients and 42 financial, behavioural, demographic and regulatory variables. The data were split into training (2020-2022), validation (2023) and out of time test (2024) sets. A rule based scheme, a logistic regression and machine learning models (Random Forest and XGBoost) were compared, and their performance was assessed using AUC ROC, KS and operational metrics at the decision threshold defined in the validation stage. In the test set, Random Forest achieved an AUC ROC of 0.87 and a KS of 0.58, and XGBoost an AUC ROC of 0.86 and a KS of 0.55, clearly outperforming the traditional rule based approach (AUC 0.59; KS 0.18) and increasing the recall of the default class from 0.66 to 0.87, with temporal stability (PSI≈0.0066) and good calibration (Brier≈0.05-0.06). The reduction in false negatives translated into avoided losses estimated at L 13 million and a ROI of approximately 825.95 %. A pilot implementation of the hybrid model is recommended under a model risk governance framework aligned with Basel III, BCBS 239 and SR 11 7.
Keywords
Aprendizaje Automático, Banca Empresarial, Honduras, Modelo Híbrido, Riesgo Crediticio