Evaluating Machine Learning Approaches for Forecasting Creditworthiness of Industrial Companies in Low-Default Portfolios

The study focuses on estimating the creditworthiness of industrial enterprises in the Eurozone and North America, using a sample of 122 companies from 2008 to 2022. The data framework consists of annual financial, macroeconomic, and ESG risk factors. The paper’s principal contribution lies in a comparative examination of advanced modelling approaches for assessing the credit quality of corporate low-default portfolios. As a LightGBM classifier reveals superior prediction quality, an original assumption that a combined model has a stronger discriminating power is debunked. Since credit rating models have comparable variance, their combination, even when supplemented with the probability of default model’s estimates, fails to correct for the observed bias, producing less accurate results than standalone machine learning algorithms. Furthermore, a complex ensemble of models appears to be impenetrable and unstable over time, whereas artificial intelligence models, such as random forest and gradient boosting classifiers augmented with the SHAP framework, are appropriate for capital calculations. Ultimately, the constructed models could be used by financial institutions and investors to evaluate the financial standing of low-default portfolios comprised of industrial firms from advanced economies.