Neural Networks, Ordered Probit Models and Multiple Discriminants. Evaluating Risk Rating Forecasts of Local Governments in Mexico

Credit risk ratings have become an important input in the process of improving transparency of public finances in local governments and also in the evaluation of credit quality of state and municipal governments in Mexico. Although rating agencies have recently been subjected to heavy criticism, credit ratings are indicators still widely used as a benchmark by analysts, regulators and banks monitoring financial performance of local governments in stable and volatile periods. In this work we compare and evaluate the performance of three forecasting methods frequently used in the literature estimating credit ratings: Artificial Neural Networks (ANN), Ordered Probit models (OP) and Multiple Discriminant Analysis (MDA). We have also compared the performance of the three methods with two models, the first one being an extended model of 34 financial predictors and a second model restricted to only six factors, accounting for more than 80% of the data variability. Although ANN provides better performance within the training sample, OP and MDA are better choices for classifications in the testing sample respectively.