A Quantitative Analysis of Default Risk Using Machine Learning and SHAP Value Interpretation

In finance, creating a model that balances risk reduction with opportunity is essential. This investigation addresses the necessity for risk evaluation frameworks that combine efficiency with adaptability, thus preserving opportunities for transactions critical to some organizations. The present study identifies, within a factoring process involving two key players: the invoice seller and the debtor, the essential variables that determine the likelihood of the debtor defaulting on the invoice payment. The event of non-payment is most often associated with the debtor's inability to pay due to insolvency, making it crucial in this type of activity to emphasize an efficient credit scoring system capable of proactively highlighting a debtor company with a high risk of default. Nonetheless, some companies pass this filter and enter the factoring process, but end up being unable to pay. The study identifies them based on a set of real data and uses supervised machine learning techniques to select the optimal classification model, also highlighting the variables with a major impact on the target. The specialized literature is focused on identifying the models that perform best in the credit scoring activity or studies that identify the non-payment behavior of clients. What this work adds is the combination of these two dimensions, for example, it provides an additional filter to credit scoring, using parameters identified as essential in determining defaulters and using them as inputs for an unsupervised learning model, thus classifying the entire population of companies in Romania to identify clusters containing the highest proportion of non-payment companies.