Credit Rating Models based on Backpropagation Neural Networks, a Peruvian Case

Microfinance institutions use various credit rating models to improve their credit risk management efficiency. These models are based on statistical and Artificial Intelligence (AI) techniques. We work with a database from a Peruvian microfinance company that, for security reasons, we only knew the description of the variables. This database contains the records of 15,569 borrowers with 26 variables for the classification of their clients and a final variable (27) for acceptance or rejection of credit. In the statistical part, an exhaustive study of the interdependence of the variables is carried out, discovering that the variable (27) depends on a single decision-making variable (8). The influence of the other variables on the decision-making variable is analyzed. In the (AI) part, various Backpropagation Neural Network (BPNN) structures are tested, taking the complete database and obtaining good precision in the prediction, reaching 97.68%. This high precision of the neural network is explained because the final variable (27) depends directly on the decision variable (8). With the same BPNN structures and taking variable (8) instead of variable (27), the precision obtained by the neural network decreases to 77.40%. Finally, maintaining the variable (27) as the expected value of the network and eliminating the variable (8) from the database, the precision of the neural network drops to 66.89%, which confirms that this variable is the most realistic.