Look Who’s Talking: Interpretable Machine Learning for Assessing Italian SMEs credit default

Small and Medium Enterprises (SMEs) are the backbone of the European Union (EU). They account for the majority of the value-added within the non-financial sector and employ two-thirds of the workforce. Therefore the assessment of the SMEs' creditworthiness is a significant issue within both financial and non-financial organizations. Recent works suggest that SMEs' default prediction is more complex than in large enterprises, and canonical models may fall short. At the same time, scholars pointed out similarities with the retail credit risk industry. In this context, black-box machine learning models are the standard, and they achieved remarkable performances. However, these models are less frequently used in enterprise credit scoring as they are essentially uninterpretable. This work aims to model SMEs' defaults using uninterpretable black-box models and compare them with the canonical models. We first explore these models' effectiveness for default prediction compared to standard models (Logistic, Probit, and Binary Generalized Extreme Value). Second, we overcome the lack of interpretability by using recent model-agnostic techniques such as Accumulated Local Effects and Shapley values. We use the dataset to estimate the models and predictions containing Italian Small and Medium Enterprises' information belonging to the manufacturing sector. The results suggest that the black-box models outperform the standard models used as the benchmark.