Evaluation of Neural Network and Logit Models for Classification of Default in Banking Loans

The purpose of the study was to evaluate the performance of neural networks as modern techniques to classify the risk of default against the traditional Logit statistical method, taking a Honduran bank as a case study. The data was obtained from its credit portfolio made up of 38,156 personal loans and 9 available characteristics, choosing the most representative independent variables to design a Multilayer Perceptron type base model and its Logit equivalent to which characteristics were added to analyze their impact on the classification of the dependent variable Default, leaving in the end a network with an input layer of 8 nodes, 4 hidden dense layers of 20 and 24 nodes, a central dropout layer and a node in the output layer as well as an equivalent logistic regression to compare the performance of both. The results with unbalanced data showed a superior performance of the networks, but when applying SMOTE oversampling, although there was no greater impact on the network, there was in the regressions, concluding that these learn to classify loan default better when the data subsets are balanced in the class of the response variable since its new results almost reached those of the neural network, which was finally chosen as the preferred model for its implementation with an accuracy of 99.16%, precision of 99.47%, sensitivity of 99.59%, specificity of 95.48 %, F1 score of 99.53% and ROC and PR curves with AUC of 98.68% and 97.69% respectively.