The Tale of Two Techniques - The comparative accuracy of machine learning and statistical techniques in predicting corporate bankruptcy for Swedish industrial firms

Abstract: Bankruptcy prediction has long been an important area of study, yet the evolution of these predictive models in the context of modern machine learning techniques remains underexplored. Our thesis addresses this by comparing the effectiveness of probit analysis - a time-tested statistical approach - with XGBoost - a new-era machine learning technique - in predicting corporate bankruptcy among Swedish firms. Utilizing a dataset of Swedish industrial firms, we meticulously assess the accuracy of each model, looking also at their capacity to select and leverage relevant independent variables. Our findings reveal notable differences in the performance of these models, providing valuable insights for researchers and practitioners. While the probit model offers a reliable, well-established framework, XGBoost demonstrates superior adaptability and performance, marking a significant advancement in bankruptcy prediction methodologies. The machine learning technique also proves better at extracting useful information through feature selection and appears more generalizable when tested on firms of different industries and sizes. We perform several robustness checks to ensure the viability of these conclusions and end by discussing our findings, limitations and potential future research directions.