대안 데이터를 활용한 신용 평가: 빅데이터와 머신러닝 활용 연구 = Credit Evaluation Using Alternative Data: A Study on the Utilization of Big Data and Machine Learning

국문 초록 (Abstract)

본 논문은 신용 평가 시스템에서 대안 데이터를 활용하여 금융 이력이 부족한 사람들(씬파일러)의 신용 평가를 개선하는 방법을 제시한다. 대안 데이터는 전통적인 금융 정보가 부족한 경우에도 추가적인 정보를 제공하여 신용 평가의 정확성을 높이고, 금융 포용성을 확대하는 데 기여할 수 있다. 이를 위해 2020년 5월부터 10월까지 카카오뱅크 대출 신청자 3,311,744명의 금융 및 대안 데이터를 분석하였으며, 머신러닝 기법인 XGBoost 알고리즘을 적용하여 기존 신용평가 모델보다 우수한 성능을 가진 대안 신용 평가 모델을 개발하였다. 본 연구의 대안 신용 평가모형은 AUC 87%의 예측력을 기록하며, 동일 대상자에 대해 산출된 한국 주요 신용평가사의 표준신용점수(AUC 84%, 82%)보다 높은 성능을 보였다. 특히, 금융 소외 계층, 비급여 근로자, 젊은층 등 기존 신용 기록이 부족한 집단에 대한 공정한 평가를 가능하게 했다. 본 연구는 대안 데이터가 신용 평가의 정확성을 높일 뿐만 아니라, 금융 포용성을 증대시킬 수 있음을 시사하며, 이러한 데이터 활용 시 개인정보 보호 규정 준수의 중요성도 함께 강조한다.

다국어 초록 (Multilingual Abstract)

In this study, we developed a two-layer credit evaluation model using the XGBoost algorithm, leveraging a wealth of alternative data from various sources. Our primary objective was to evaluate the effectiveness of alternative information in predicting the delinquency of loan applicants. This assessment involved comparative analyses between scores generated from alternative data sources and credit scores established by credit bureaus.
This study has explored the largest number of alternative information sources among domestic empirical studies. Unlike previous studies that primarily focused on the usefulness of alternative information as a complementary factor to credit information, this study has demonstrated that alternative information alone can also provide excellent performance when a diverse range of alternative information is used.
The model we developed produced an alternative score with higher predictive power, achieving an AUC of 87%, compared to the standard credit scores from credit bureaus, which had AUCs of 84% and 82% respectively. This demonstrates that the alternative score not only outperforms standard credit scores but also exhibits a high level of robustness.
A key finding of our research is that by utilizing a sufficient amount of alternative data, it is possible to achieve higher predictive accuracy than standard credit scores, even without relying on demographic variables. This enables a fairer evaluation of loan applicants, particularly for financially marginalized groups such as non-salaried workers and younger individuals who lack traditional credit histories.
Furthermore, contrary to concerns that the utilization of alternative data in credit scoring may exacerbate social stratification, this study demonstrates that alternative data can provide significant benefits to those who have been disadvantaged by the traditional credit scoring system, which has primarily focused on salaried individuals and may have led to conservative evaluations for other groups. This suggests that alternative data not only enhances the accuracy of credit evaluations but also increases financial inclusion.
Additionally, our research underscores the importance of adhering to data privacy regulations when utilizing alternative information, as the use of alternative data becomes more prevalent in credit scoring.
Overall, this study highlights the potential of alternative data to improve the credit scoring system by enhancing accuracy and inclusivity. The findings suggest that integrating alternative information into credit evaluation models can provide considerable benefits, both in terms of predictive performance and in promoting financial inclusion for underserved populations, especially thin-filers.