The Korea Capital Market Institute has developed a machine learning model capable of detecting signs of corporate insolvency before delisting. This study aimed to analyze the financial statements of listed companies in Korea to identify early signs of financial distress. Consequently, it is expected to help domestic investors and financial authorities assess corporate financial soundness more promptly.
The research team at the Korea Capital Market Institute incorporated the growing significance of footnote disclosures under the Korean International Financial Reporting Standards (K-IFRS). They developed a multimodal neural network-based machine learning model that not only utilizes quantitative financial statement data but also learns from unstructured information contained in footnotes. According to the researchers, this model enables a more sophisticated detection of financial distress compared to conventional analysis methods.
The research team trained the model using data from 16,815 firm-year observations of non-financial firms listed on the KOSPI and KOSDAQ markets with December fiscal year-ends between 2005 and 2019. The model’s predictive power was tested by analyzing 43 quantitative financial indicators, 37 key account items, and all unstructured information contained in financial statement footnotes. The results confirmed that the model could accurately predict delisting risks before trading suspension occurs.
In particular, the research team highlighted the prevalence of delisting following trading suspension in the Korean market. According to the Korea Exchange’s regulations, once a company initiates the delisting process, its stock may remain suspended from trading for an extended period, making it difficult for investors to respond appropriately. To address this issue, the team successfully developed a machine learning model capable of detecting financial distress before trading suspension, establishing an effective early warning system.
However, the researchers also pointed out the limitations of footnote disclosures in predicting financial distress and emphasized the need for further improvements. Despite the vast amount of information provided in footnotes, their contribution to predicting insolvency was found to be limited. This finding suggests the need for a thorough review of whether the principle of accounting conservatism is adequately reflected in footnote disclosures. The team recommended strengthening the link between financial statement footnotes and main financial data while enhancing the clarity and comparability of disclosures.
This study holds significant implications, demonstrating that a machine learning-based insolvency prediction model can serve as a practical risk management tool in the Korean financial market. Moving forward, the research team plans to enhance the quality of footnote disclosures to improve the accuracy of the predictive model further. Through these efforts, the study is expected to contribute to investor protection and financial market stability in Korea.
