AI credit scoring reads business networks, CIKM paper accepted

If the closure of a single business partner is a risk signal that begins outside the books, the AI developed by Douzone Bizon, Techfin Ratings and the Korea Advanced Institute of Science and Technology (KAIST) follows how far that signal spreads. The paper on the business-relationship network credit-scoring model ‘DefaultGNN,’ jointly developed by the three organizations, was accepted at the international CIKM conference. CIKM is an international conference covering artificial intelligence, data mining and information retrieval.

Traditional corporate credit assessments have centered on individual companies’ financial information, including profitability, stability, liquidity and growth. DefaultGNN adds network variables such as business partners, transaction categories and shares, business suspensions and closures, defaults, and canceled amounts. Graph neural networks (GNNs)—an AI technology that learns the connection structures between companies—are used to read how risk propagates through business networks.

The researchers confirmed that defaults and business suspensions among second- and third-tier partners also have a cascading impact on the company’s default rate, in addition to those of direct first-tier partners. They explained that combining the model with existing models improved credit-prediction performance over the standalone model and made predictions more precise even for companies with limited financial statements. The results also showed that corporations achieved higher loan approval rates while keeping default rates stable, while sole proprietors saw approval rates rise and default rates fall at the same time.

The groups most affected by this change would be small and midsize companies and sole proprietors that were assessed more conservatively than their actual business conditions because they lacked sufficient financial statements. If healthy relationships with business partners and continued business activity are reflected in credit decisions, financial institutions could gain a way to consider repayment capacity that cannot be seen from the books alone. However, this is only a possibility presented by the research. Acceptance of a CIKM paper does not mean deployment as a commercial financial service.

In this study, Douzone Bizon’s enterprise resource planning (ERP) data became more than a simple record of transactions: it served as material for reading the context of relationships. The company said it combined decades of accumulated ERP data with AI algorithms. The findings point not to abandoning financial-statement-centered assessments, but to supplementing them with information from external business networks. The work is still at the research stage, but it marks a starting point for enabling finance to read not only a company’s numbers but also the ecosystem to which it is connected.