El Modelo Logit Mixto para la construcción de un Scoring de Crédito
Cargando...
Autores
Director
Tipo de contenido
Document language:
Español
Fecha
Título de la revista
ISSN de la revista
Título del volumen
Documentos PDF
Resumen
Resumen: En este trabajo se presenta un modelo logit mixto aplicado a datos reales para un scoring de otorgamiento de créditos, como una propuesta alternativa para evaluar los supuestos con respecto a la estructura de covarianza (independencia y homocedasticidad) que asumen los modelos logit y probit, tradicionalmente utilizados para medir el riesgo crediticio. El modelo logit mixto se evalúa en contraste con el logit y el probit, partiendo desde el enfoque de la teoría económica de la utilidad aleatoria, con respecto a la decisión que enfrenta el cliente de pagar o no el crédito, según su función de maximización de utilidad. De acuerdo con las medidas de bondad del ajuste (sensibilidad y especificidad), el modelo propuesto (logit mixto) es más sensible que los tradicionalmente usados (logit y probit), puesto que detecta en mayor proporción los clientes que entran en default. Sin embargo, aunque los tres modelos ajustados tienen un alto poder discriminatorio (AUROC), el logit mixto presenta la menor tasa de especificidad.
Abstract: In this thesis we present a mixed logit model applied to real data for a credit scoring, which is proposed as an alternative approach to evaluate the assumptions about the covariance structure (independence and homoscedasticity) to the logit and probit models, traditionally used to measure credit risk. The mixed logit model is evaluated in contrast to logit and probit models. We start from the perspective of the economic theory of a random utility model that describes the decision faced by the individual to pay the credit, after maximizing his/her utility function. According to the goodness of fit measures (sensitivity and specificity), the proposed model (mixed logit) is more sensitive than the traditional models (logit and probit), as it was detected in the greater proportion of customers in default. Even though the three fitted models have a high discriminatory power (AUROC), the mixed logit has the lowest rate of specificity.
Abstract: In this thesis we present a mixed logit model applied to real data for a credit scoring, which is proposed as an alternative approach to evaluate the assumptions about the covariance structure (independence and homoscedasticity) to the logit and probit models, traditionally used to measure credit risk. The mixed logit model is evaluated in contrast to logit and probit models. We start from the perspective of the economic theory of a random utility model that describes the decision faced by the individual to pay the credit, after maximizing his/her utility function. According to the goodness of fit measures (sensitivity and specificity), the proposed model (mixed logit) is more sensitive than the traditional models (logit and probit), as it was detected in the greater proportion of customers in default. Even though the three fitted models have a high discriminatory power (AUROC), the mixed logit has the lowest rate of specificity.