Modelo de selección de riesgo, aplicado al otorgamiento de microcrédito en un intermediario financiero de crédito colombiano, apoyado en redes neuronales artificiales

This article aims to identify the most relevant variables that allow through a neural network model (RNA), with supervised learning, in a kind of error correction and feedforward perceptron multilayer architecture to achieve the best predictors of low risk, in the process of microcredit. Academic studies on the topic are limited, since the raw material for your application, in most cases is not made public. This information (databases), it becomes an intangible asset for financial institutions, who permanently develop methodologies for the grant, being the essence of your business, in managing risks. Also implies that, according to the experience every time record is unique and particular and universal cannot make the application of a model. Information was available on the behavior of one hundred and sixteen thousand borrowers in a horizon of three years, after debugging. The study found that the selected model (M105), provides a better estimate of borrowers classified as "Hazardous" with a 38.8% and an overall 58.5% correct. In the case of borrowers 45,953 presented to the network as "Risky", failed to qualify in this category to 17,853. Also there is a wrong classification rate of 61.1%, corresponding to 28,100 borrowers. That is, the number of borrowers classified as "Not Hazardous".