Bankruptcy Prediction of Danish SMEs Using Machine Learning Models: A Comparative Analysis of Predictive Models Across Industries

Abstract

The substantial societal implications of the financial collapse of companies have historically been a widely investigated subject of academic research, fueling collective motivation to understand and predict the event and consequences thereof. Recent developments have increased the sophistication of prediction methods, facilitating an increased need for empirical testing. This thesis investigates the predictive capabilities of various machine learning models, using financial data collected from Danish small and medium enterprises, exploring the research question: “How effectively can the applied machine learning models predict bankruptcy among Danish SMEs? The methodologies were employed to enable the optimal processing of models by using a wide selection of techniques to collect, parse, preprocess, regulate, and assess the data, models, and results. The applied models of the thesis were Logistic regression, Decision tree, Random Forest, Support Vector Machine, and eXtreme Gradient Boosting. The data was subset into thirteen different Industries. Three different iterations of the models were created, differentiated by the underlying data: General Models, Production- vs. Service Industry Models, and Industry-specific models. The results indicated Random Forest as the overall best-performing model, reaching Aucscores of 0.98. The predictive performance of the models was notably higher on the General and Production- vs. Service industry models. The models showed the best predictive performance on Industry Q (Human health and social work activities). Examining the results of the Random Forest model, the features with the highest contribution to impurity reduction were based on financial ratios that express Liquidity, Capital Structure, and Leverage, indicating Profitability-based ratios as the least important features.