Interpretable selective learning in credit risk

Forecasting credit default risk has been an important research field for several decades. Traditionally, logistic regression has been widely recognized as a solution because of its accuracy and interpretability. Although complex machine learning models may improve accuracy over simple logistic regressions, their interpretability has prevented their use in credit risk assessment. We introduce a neural network with a selective option to increase interpretability by distinguishing whether linear models can explain the dataset. Our methods are tested on two datasets: 25,000 samples from the Taiwan payment system collected in October 2005 and 250,000 samples from the 2011 Kaggle competition. We find that, for most of samples, logistic regression will be sufficient, with reasonable accuracy; meanwhile, for some specific data portions, a shallow neural network model leads to much better accuracy without significantly sacrificing interpretability.