Corporate insolvency, particularly in SMEs, poses significant economic risks due to limited funding options and potential supply chain disruptions. Traditional insolvency prediction models relying on financial information face limitations when applied to SMEs due to insufficient accounting practices, simplistic business structures, unreliable financial data, and the annual nature of financial statements. This study addresses these challenges by utilizing non-financial data—specifically, technological feasibility assessment data from the Korea SMEs and Startups Agency (KOSME)—to create a more robust and easily interpretable insolvency prediction model for SMEs. The study focuses on developing an insolvency prediction model using a decision tree algorithm, known for its explainability, and then simplifies the model's output through a novel rule selection method, the HSC, making it easier to understand and act upon the factors leading to insolvency in SMEs. This is critical for providing targeted, practical insolvency prevention strategies tailored to the specific characteristics of different SME types.

Existing literature extensively uses financial ratios (e.g., Altman's Z-score) for insolvency prediction, but these methods have limitations with SMEs. Recent research increasingly incorporates non-financial information, demonstrating improved prediction accuracy when combining financial and non-financial data. Studies using technological feasibility assessments as a non-financial indicator show promise in predicting insolvency, particularly for SMEs with limited reliable financial data. This study builds upon this research by focusing on the use of technological feasibility assessments and developing a method to simplify the interpretation of decision tree models. The review also covers existing rule selection methods from decision trees, such as Laplace accuracy and Weighted Relative Accuracy (WRA), which serve as a basis of comparison for the proposed HSC method.

The study uses a five-stage research framework. First, data on technological feasibility assessments from KOSME were collected and preprocessed. SMEs were categorized into three types: general, technology development, and toll processing. To address data imbalance (more healthy than insolvent companies), various data balancing techniques were applied: under-sampling (random under-sampling, SpreadSubsample, ClusterCentroids) and over-sampling (random over-sampling, SMOTE, ADASYN). Feature selection using backward elimination with gain ratio as a criterion was employed to identify crucial variables. A decision tree algorithm was then used to develop three insolvency prediction models, one for each SME type. The HSC method, a novel rule selection technique, was then applied to the decision trees. The HSC method combines support and confidence measures to identify the most important rules. This selection process aims to reduce the complexity of the decision tree while retaining its predictive power, resulting in a CorDT for each SME type. These CorDTs formed the basis for explaining the causes of insolvency and developing customized prevention strategies for each SME type. A cost-sensitive approach was also incorporated during model building to address the imbalance in class distribution in the datasets.

The study found that the SMOTE over-sampling technique yielded the highest average hit ratio (77.6%) across the three SME types in the insolvency prediction models. The cost-sensitive approach did not significantly improve performance. The HSC method, compared to Laplace and WRA, showed a more balanced consideration of classification accuracy and the number of observations, resulting in a more effective selection of important rules. Analysis of the CorDTs revealed distinct insolvency factors for each SME type. For general type SMEs, credit status and management stability were most crucial. Technology development type SMEs showed competitive strength and sales management as key factors. Toll processing type SMEs were most affected by business propulsion, management stability, and CEO reliability. The CorDT models, consisting of a selected subset of the most important rules based on HSC, were highly effective in explaining the majority of insolvency cases for each type.

The findings highlight the importance of using non-financial data, such as technological feasibility assessments, to predict insolvency in SMEs where financial data might be unreliable or incomplete. The successful application of SMOTE demonstrates the effectiveness of addressing data imbalance in improving prediction accuracy. The development of the CorDTs, using the proposed HSC method, provides easily interpretable models that offer actionable insights for insolvency prevention. The type-specific insolvency factors identified provide targeted recommendations for SMEs, enabling more effective resource allocation and risk mitigation strategies. The research contributes to the field by offering a new method for generating simplified yet accurate insolvency prediction models suitable for SMEs with limitations in readily available financial data.

This study successfully developed a decision tree-based insolvency prediction model for SMEs using non-financial data and the novel HSC rule selection method. The CorDTs provide a simplified, interpretable model for each SME type, highlighting crucial insolvency factors and suggesting tailored prevention strategies. Future research could incorporate macroeconomic indicators, further enhance the HSC method, and explore the application of anomaly detection techniques to improve model robustness.

The study's scope is limited to SMEs in the manufacturing sector in Korea, using data from a single year. The reliance on self-reported data from KOSME assessments might introduce some bias. The study does not fully account for external environmental factors beyond the scope of the technological feasibility assessment, like macro-economic conditions and the impact of broader economic downturns.