SCORING CARDS FOR DIFFERENT TYPES OF CREDIT PRODUCTS

The work takes into account contemporary techniques for creating scoring models for different kinds of credit products using machine learning techniques and alternative data sources. The models (logistic regression, Random Forest, XGBoost, and LSTM) were compared for the first time for Kazakhstan's second-level banks, with a focus on their interpretability, long-term stability, and suitability for various lending markets. The study's scientific novelty lies in how it uses Explainable AI techniques (SHAP, PDP) and alternative data (digital behavior, transactions, and telecom data) to create scoring cards and assess temporal stability models using the population stability index (PSI). The findings indicate that XGBoost exhibits the highest accuracy (AUC = 0.88), Random Forest offers a balance between quality and interpretability (AUC = 0.82), and logistic regression continues to be the fundamental model for regulated products because of its transparency and adherence to regulations