머신러닝을 이용한 창업기업의 대출부도 예측

Although machine learning(hereafter ML) has been actively applied to recent research on predicting the default of small and medium-sized enterprises(hereafter SMEs), studies using ML to predict loan defaults of startups are rare. To bridge the gap in the field of SMEs default prediction research, this study aims to evaluate the accuracy of ML for predicting loan defaults of startups. To this end, we compares which ML model is most accurate for predicting loan defaults of domestic startups and established companies, and how its performance differs from statistical models. The main findings from the empirical results are summarized as follows: Firstly, AdaBoost has been evaluated as highly accurate and stable for predicting short- to long-term (1 to 3 years) loan defaults in startups, while support vector machines showed good performance for 1-year predictions in established company. Secondly, the accuracy of loan default prediction for startups using statistical models, such as the logistic regression, was found to be nearly equivalent to that of ML. However, there were ML techniques that performed worse than statistical models. Thirdly, financial features provide significant explanatory power in predicting loan defaults of startups. Debt capacity ratio and liquidity ratios were effective for both short- and long-term loan default predictions in startups and established businesses. The analysis suggests that while machine learning significantly improves the accuracy of predicting loan default for SMEs, it has the drawback of lower interpretability of predictive variables. Therefore, its usefulness varies depending on the purpose of use. This paper contributes by evaluating the usefulness of multiple ML and the effectiveness of financial features for predicting the loan default of startups through an intertemporal validation method.