CMStatistics 2021: Start Registration
View Submission - CMStatistics
Title: Black-box models and Interpretability: Prediction of Italian SMEs' default Authors:  Lisa Crosato - Ca Foscari University of Venice (Italy)
Caterina Liberati - University of Milano-Bicocca (Italy)
Marco Repetto - University of Milano-Bicocca (Italy) [presenting]
Abstract: Assessing Small and Medium Enterprises (SMEs) creditworthiness is a significant issue within organizations. Recent works suggest that SMEs' default prediction is more complex than large enterprises. To handle this complexity, both scholars and practitioners resorted to model credit risk through Machine Learning (ML) techniques, whose applications saw a steep increase in the retail credit risk ambit. However, the lack of interpretability of black-box models has limited their usage in credit risk applications. A possibility of restoring interpretability can be found in reverse-engineering the ML model without accessing its inner parameters, leading to post-hoc explanations that allow the Decision Maker to pin down the relevant effects captured by the black-boxes and decide accordingly. The aim is to model and interpret SMEs' defaults using the eXtreme Gradient Boosting (XGBoost) and the FeedForward Neural Network (FANN) algorithms, and compare them with a few traditional models. We employ recent model-agnostic techniques, such as Accumulated Local Effects and Shapley values, to overcome the usual lack of interpretability of the black-box machines. Results show an overall highest classification power by the XGBoost and highlight the ranking of the input variables based on their contribution to the model outcome as well as the variables impact on the likelihood of default.