Analytics

In recent years, the micro-lending and retail banking sectors have seen a growing trend where some borrowers initiate bankruptcy proceedings shortly after obtaining a loan. For the lender, this results in a sharp deterioration of expected cash flows from issued loans, an increase in the share of bad debt, and additional costs associated with managing distressed debt and legal proceedings. At the portfolio level, even a moderate rise in the frequency of such events can significantly worsen the risk profile and financial performance, particularly for high-turnover products with short lending horizons.

       Traditional scoring models—which primarily focus on historical delinquency and current debt burden—may fail to identify early enough those clients whose behavior signals an intention or high probability of future bankruptcy. Consequently, there is a need for a tool capable of assessing the likelihood of a borrower declaring bankruptcy in the near future as a distinct target event during the initial credit decision process.

       Unlike the standard prediction of delinquency, predicting the probability of bankruptcy requires analyzing not only financial insolvency but also credit behavior patterns preceding the legal event. These include the accumulation of liabilities, the frequency of credit inquiries and new loan originations, payment discipline stability (or instability), the frequency of loan restructurings, and indicators of over-indebtedness and debt concentration. Experience shows that potential bankrupts often exhibit specific indirect characteristics in their credit history; these can be formalized as model features. The focus here is not on isolated "red flags," but rather on a combination of signals and their evolution over time, enabling the distinction between borrowers facing temporary financial difficulties and clients with a high probability of proceeding to formal bankruptcy. The following section describes the practical component of the study, focusing on the construction of a model to predict the occurrence of legal bankruptcy, as well as the principles for creating the training dataset and defining target functions. The results obtained can be utilized by risk analysts and product owners to improve scoring quality and reduce losses associated with subsequent borrower bankruptcy.

       The initial dataset comprised approximately 16,000 borrowers whose credit histories included records of bankruptcy proceedings—whether ongoing or concluded (no distinction was made regarding the specific bankruptcy status during selection). This approach ensured a sufficient volume of observations unequivocally confirming the occurrence of the target event. Subsequently, historical applications from some of these borrowers—dating back to the period prior to the appearance of bankruptcy indicators in their credit histories—were identified within the accumulated application database. In other words, at the time these applications were submitted, the borrowers did not yet have a recorded bankruptcy proceeding, although such a proceeding occurred later. These observations were assigned to the "bad" class within the modeling dataset.

       To form a control group of "good" observations, applications were randomly selected from comparable time periods for which no bankruptcy proceedings were subsequently recorded. Selecting from similar timeframes is crucial for mitigating bias risks associated with changes over time in macroeconomic conditions, credit policies, and the accessibility of bankruptcy procedures.

Based on the resulting dataset, binary target variables were defined to reflect the occurrence of bankruptcy across various forecasting horizons. Specifically, the following target variables were used: "bankruptcy within 30 days," "within 60 days," "within 90 days," and "within 180 days," as well as interval-based targets defined as "between 30 and 90 days" and "between 90 and 180 days." This set of target functions allows for comparing prediction quality across both short horizons (which are more sensitive to immediate risks) and longer ones. Credit history characteristics available at the time of the application decision served as predictors.

       Modeling was performed using logistic regression—a fundamental, interpretable, and widely used tool in credit scoring. The training process yielded models with class separation quality (AUC) exceeding 0.8, indicating high predictive power within the defined problem scope and target variables used.

       In the final stage, the models were evaluated on a test dataset; the most stable solutions were selected from the pool of candidates, and a final ensemble was constructed to enhance performance stability and reduce sensitivity to the specific characteristics of individual targets and data segments. The performance metrics of the final ensemble across the selected target variables are as follows:

       A key outcome is the model's high practical value for underwriting and scoring processes. Implementing the model allows for the targeted exclusion of applications with a high estimated probability (over 70%) of initiating bankruptcy proceedings in the near future, while the impact on the portfolio's overall approval rate remains minimal. This effect is achieved due to the high concentration of future bankruptcies within a relatively small segment of applications.

       The study confirms that a borrower's legal bankruptcy is a distinct risk event requiring separate forecasting, rather than merely a "special case" of standard delinquency. The resulting models demonstrated that significant predictors of bankruptcy differ in their economic logic from predictors used in delinquency models. While Probability of Default (PD) models typically capture a deterioration in current payment discipline, the bankruptcy model relies more heavily on indirect credit history characteristics—reflecting specific behavioral patterns and structural changes in debt load prior to the onset of legal bankruptcy.

       Thus, the proposed approach provides an effective tool for minimizing loan originations to borrowers with a high probability of subsequent legal bankruptcy. Its implementation reduces expected losses and legal costs, improves portfolio resilience, and enhances risk management quality without significantly compromising business performance. Future development of the solution may include expanding the range of data sources, regularly monitoring predictor stability, and fine-tuning scoring thresholds and usage strategies to align with specific product segments and the lender's risk appetite.