머신러닝을 활용한 금융 사기기업 조기탐지 모델링 연구: 기술보증기업 데이터 기반 탐지모형 개발 및 알고리즘 성능 비교

This study develops an AI-based detection model to distinguish between legitimate firms and fraudulentfirms using data from the Korea Technology Finance Corporation (KOTEC), and compares theperformance of various machine learning algorithms. Recently, several cases have emerged in whichfirms obtained guarantees from KIBO by submitting falsified documents and subsequently committedlarge-scale loan fraud, highlighting the need for an early warning system capable of detecting fraudulentactivities at the guarantee screening stage. To address this issue, data were collected from 55,219 smalland medium-sized enterprises (SMEs) that received guarantees between 2019 and 2023. Among them,1,222 firms were defined as fraudulent firms. if they either defaulted within one year after receivinga guarantee or were detected and investigated by the police for fraud. A total of 59 candidate variables—including firm characteristics, representative attributes, financial indicators, and technology-relatedmeasures—were considered, from which 18 significant predictors were selected through statisticaltesting and feature importance analysis. Preprocessing steps such as outlier removal and Min-Maxscaling were applied. To address the severe class imbalance, Propensity Score Matching (PSM) wasemployed to construct a balanced training dataset consisting of 1,222 fraudulent firms and 1,222non-fraudulent firms. Multiple algorithms, including logistic regression, random forest, XGBoost, andartificial neural networks, were trained, with hyperparameters optimized via cross-validation. Theempirical results using an independent test set revealed that the random forest model achieved thehighest recall and F1 score, demonstrating superior detection capability for fraudulent firms. Logistic regression, although yielding relatively low recall, showed the best classification power in terms ofROC-AUC and could be practically useful when classification thresholds are adjusted to improvesensitivity, despite the resulting decline in precision. XGBoost and neural networks exhibitedcomparatively lower performance in both recall and F1 score, with neural networks showing the weakestrecall, likely due to insufficient generalization under limited training data. Nevertheless, logisticregression offered advantages in terms of probabilistic consistency and interpretability. Based on itspredictive probabilities, a ten-tier risk grading system (AAA to D) was designed, which stratifies firmsinto percentile-based risk categories: the top 10% (AAA) as the safest and the bottom 10% (D) asthe riskiest. This study demonstrates the practical applicability of machine learning-based early detectionmodels for fraudulent firms and proposes an interpretable grading framework that can be utilized inpolicy and guarantee screening practices. Incorporating such models into the policy financing processis expected to contribute to the efficient allocation of limited financial resources to genuine startupswhile preventing fraudulent loan practices.