Análise de um sistema de crédito cooperativo através de redes neurais (MLP) com a utilizaçao do algoritmo Levenberg Marquardt
Resumo
Resumo: Os problemas envolvendo tomada de decisões estão presentes em todas as áreas da atividade humana e, ao longo da história, foram tratados através de uma série de abordagens diferentes. Este estudo propõe uma abordagem estruturada na exploração das Redes Neurais Artificiais (RNA's) como instrumento de apoio à previsão de inadimplência de pessoa física, utilizando o banco de dados de uma Cooperativa de Crédito da região do sudoeste do Paraná, que contém informações de seus associados/clientes. Face a necessidade de se tomar decisões bem estruturadas e com base no histórico de pessoas físicas, o modelo de avaliação do risco de inadimplência foi construído usando dados disponíveis da cooperativa de crédito, a qual possui tanto informações sobre os associados adimplentes, quanto dos inadimplentes. A metodologia do estudo propõe inicialmente, uma avaliação estatística dos dados tanto pela análise de componentes principais quanto pela análise fatorial, que teve o propósito de avaliar a importância relativa das diferentes variáveis envolvidas no processo de análise de crédito. As redes neurais desenvolvidas foram do tipo feed-forward, totalmente conectadas. Para se determinar os "melhores" pesos de treinamento foi utilizado o algoritmo de Levenberg Marquardt. Finalmente, o sucesso da modelagem foi avaliado através de uma amostra de controle, utilizada para testar a capacidade de classificação das redes neurais, sendo esta amostra diferente da usada para obter o modelo que decidirá a respeito da concessão (ou não) do crédito ao associado/cliente. Abstract: The problems that evolve how to get decisions are present in all the areas and, along the history, they were treated through a series of different discussions. This application suggest an discussion that have structure in the exploration of the "artificial neural network (ANN)" like an instrument to give support to the estimate of the risks caused by the fisical person adimplence, using the information bank of a Credit Co-operative in the South West Paraná region. That has information about it associates/clients. The obtainment of the information, that compose an offer of a credit concession, has to be versatile and assorted, composite by information that are important in the process of credit analysis, what means that it consent evaluate the grade of the risk that it have in the moment that it gives credit to the client. In face of the necessity to take decisions with good structures and based in the description of the fisical person, the model of the adimplence risk estimate will be constructed using informations desponible by the credit co-operative, that have information about the adimplentes and inadimplentes associates. It will be selected variable through the importance of wich one in the inadimplence risk estimate to the training of the net. Finally, the success of the modality will be appraised throught an control sample, that will be used to test the capacity of classification of the neural network, beinz this sample different of the another one used to get the model that will decide the concession or not of the credit to the associate/client.
Collections
- Teses & Dissertações [9329]