| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 1.94 MB | Adobe PDF |
Authors
Abstract(s)
A presente dissertação apresenta os resultados de uma pesquisa sobre métodos estatísticos mais adequados que os bancos portugueses podem usar para validar os modelos internos de cálculo de probabilidade de incumprimento de clientes de crédito.
Com a implementação do Acordo de Basileia II por volta dos anos 2007 a 2009, os bancos portugueses passaram a ter autonomia para estimar, usando metodologias internas, a probabilidade de incumprimento dos seus clientes. Com essa permissão, veio a obrigação de validar tais metodologias usando algumas técnicas estatísticas como: Curva ROC, índice AUC, Pietra Index, Curva CAP Accuracy Ratio, Matriz de confusão ou BRIER score, sendo que a validação deve ser feita por um departamento interno e independente dentro do banco. Neste sentido, o principal objetivo do presente trabalho é analisar a adequabilidade das técnicas estatísticas usadas pelos bancos para validar os modelos internos de classificação de risco de crédito (cálculo da probabilidade de incumprimento) de modo a aferir, numa base comparativa, sobre a capacidade de validação e qual das técnicas estatísticas acima apresentadas valida de forma mais adequada os modelos desenvolvidos.
Para atingir este objetivo, no presente trabalho, foram desenvolvidos três (3) modelos de regressão logística distintos para calcular a probabilidade de incumprimento na amostra selecionada, composta por empresas portuguesas não financeiras (base de dados disponibilizada pelo Banco de Portugal Microdata Research Laboratory - BPLIM). Os resultados mostraram que a regressão logística é um modelo adequado para a classificação dos clientes no que diz respeito ao risco de crédito. Quanto às técnicas estatísticas da Curva ROC, índice AUC, Pietra Index, Curva CAP Accuracy Ratio, Matriz de confusão e BRIER score, usadas para a validação dos três (3) modelos, nenhuma delas se mostrou mais importante em detrimento da outra, pois avaliam aspetos diferentes nos modelos, tendo-se verificado uma complementaridade entre os mesmos.
This dissertation presents the results of a research on the most adequate statistical methods that Portuguese banks can use to validate their internal models for calculating the probability of default of credit customers. With the implementation of the Basel II Accord around the years 2007 to 2009, Portuguese banks were given autonomy to estimate, using internal methodologies, the probability of default of their customers. With this permission, came the obligation to validate such methodologies using some statistical techniques such as: ROC Curve, AUC index, Pietra Index, CAP Accuracy Ratio Curve, Confusion Matrix or BRIER score, and the validation should be done by an internal and independent department within the bank. In this sense, the main objective of this work is to analyse the suitability of the statistical techniques used by banks to validate the internal models of credit risk classification (probability of default calculation) in order to assess, on a comparative basis, the validation capacity and which of the statistical techniques presented above more adequately validates the models developed. To achieve this goal, in this work, three (3) distinct logistic regression models were developed to calculate the probability of default in the selected sample, composed of Portuguese non-financial corporations (database provided by the Banco de Portugal Microdata Research Laboratory - BPLIM), the results showed that logistic regression is an adequate model for the classification of customers with respect to credit risk. As for the statistical techniques of ROC Curve, AUC index, Pietra Index, CAP Curve Accuracy Ratio, Confusion Matrix and BRIER score, used for the validation of the three (3) models, none of them proved to be more important in detriment of the other, as they assess different aspects in the models, and a complementarity between them was verified.
This dissertation presents the results of a research on the most adequate statistical methods that Portuguese banks can use to validate their internal models for calculating the probability of default of credit customers. With the implementation of the Basel II Accord around the years 2007 to 2009, Portuguese banks were given autonomy to estimate, using internal methodologies, the probability of default of their customers. With this permission, came the obligation to validate such methodologies using some statistical techniques such as: ROC Curve, AUC index, Pietra Index, CAP Accuracy Ratio Curve, Confusion Matrix or BRIER score, and the validation should be done by an internal and independent department within the bank. In this sense, the main objective of this work is to analyse the suitability of the statistical techniques used by banks to validate the internal models of credit risk classification (probability of default calculation) in order to assess, on a comparative basis, the validation capacity and which of the statistical techniques presented above more adequately validates the models developed. To achieve this goal, in this work, three (3) distinct logistic regression models were developed to calculate the probability of default in the selected sample, composed of Portuguese non-financial corporations (database provided by the Banco de Portugal Microdata Research Laboratory - BPLIM), the results showed that logistic regression is an adequate model for the classification of customers with respect to credit risk. As for the statistical techniques of ROC Curve, AUC index, Pietra Index, CAP Curve Accuracy Ratio, Confusion Matrix and BRIER score, used for the validation of the three (3) models, none of them proved to be more important in detriment of the other, as they assess different aspects in the models, and a complementarity between them was verified.
Description
Keywords
Acordos de Basileia Risco de crédito Probabilidade de incumprimento técnicas internas de validação Regressão logística Basel accords Credit risk Probability of default Internal validation tech