Credit Risk Assessment with Stacked Machine Learning

The Banca d’Italia’s system for the credit assessment of non-financial firms for collateral purposes in monetary policy consists of a statistical model (S-ICAS) and of the analysts’ evaluation. We compare the performance of S-ICAS with that of artificial intelligence – machine learning (ML) – models, including deep learning. We find that deep learning improves the discriminatory power; decision tree ensembles yield a further improvement, as well as a meta-model that stacks the random forests, extreme gradient boosting, and deep learning models. We apply eXplainable Artificial Intelligence (XAI) techniques to the meta-model predictions and show that XAI can support analysts in understanding the key factors behind the differences between ML and S-ICAS predictions, thus helping refine their assessment. While interpretability issues prevent ML-based models from being a full alternative to traditional models, XAI allows for their integration within the overall credit assessment process, thus increasing its effectiveness.