AI finance company PFCT (PFC Technologies) said on Wednesday it conducted a 10-month joint study with a research team led by Professor Byung-Ro Moon (문병로) at Seoul National University’s Department of Computer Engineering to expand AI credit-scoring methodologies.

The study aimed to seek a more sophisticated AI credit-scoring model by comparing and testing various algorithms based on de-identified real financial data.

The team applied the transformer, a deep-learning architecture, to credit scoring and analyzed its characteristics and constraints. Transformers are used in areas where data order is important, such as natural language processing. The team focused on learning characteristics shown by the structure in a credit-scoring data environment that has no concept of word order.

The study found that even under the same data conditions, performance indicators and learning characteristics differed depending on the model architecture. Some tests also observed a tendency for improvement in the index (KS) that distinguishes risky customers and in recall for detecting low-credit segments compared with existing methods.

The team used the findings to review the scope and limits of applying transformer-based learning methods. The study is meaningful in that it secured comparative benchmarks and empirical data for model design and further research with a view to future practical application.

Moon said the study was meaningful in that it verified various algorithmic approaches to credit scoring in a real data environment and confirmed their potential use in financial settings. He said reviewing deep-learning-based methodologies in line with the characteristics of credit-scoring data presented standards for future practical application.

Lee Soo-hwan (이수환), PFCT’s chief executive, said the industry-academia collaboration was meaningful because it approached AI credit scoring as a research area that requires continuous expansion and validation rather than a fixed technology. He said PFCT would continue research and validation of AI credit-scoring technology based on the results combining academic algorithm research with real industry data.