A MLP based PD estimation model for SME credits

This study presents a multilayer perceptron (MLP) cascaded with a logistic regression (LR) model for the probability of default (PD) estimation of real-life Small and Medium Enterprises (SMEs). The MLP does a preprocessing for LR which is used to compute PD value. The obtained raw PD values are calibrated according to the real portfolio default average. In experiments, a Turkish SME database is used. The results indicate that the cascaded MLP-LR model provides better classification accuracy and outperforms commonly used logistic regression.