Credit Scoring Algorithm Overview
Uploaded by
Abdelaziz LotfyCredit Scoring Algorithm Overview
Uploaded by
Abdelaziz LotfyCommon questions
Powered by AIThe Kashat algorithm relies on several data types to predict repayment probabilities. These include customer phone data, historical loan attempts, and user-specific information such as financial profiling, activity indicators, travel history, and phone specifications. For new customers, the primary reliance is on phone data. In contrast, for recurring customers, the algorithm considers historical repayment behavior, which serves as a robust predictor of future repayment performance .
The Kashat algorithm faces several limitations related to its model training and deployment. The model heavily depends on the customer data it is trained on, assuming future customers will reflect a similar data profile. There's also a performance evaluation delay, as it requires 2-3 months of running time in production to analyze a model's true efficacy due to the loan cycle length .
Customers are evaluated by the Kashat algorithm at runtime, meaning their scores are generated on-the-fly as loan requests come in. This real-time evaluation aligns with the loan contract parameters, requiring a 2-3 month period to assess a model's performance post-deployment due to the loan cycle, which typically spans over two months .
The handling of new customers by the Kashat algorithm highlights their inherent risk by treating them differently due to a lack of prior loan repayment history. The algorithm uses phone data as a primary input for new customers but recognizes the higher risk associated with them. To mitigate potential losses, the business rules assign a lower initial loan limit to new customers, acknowledging the uncertainties around their repayment behavior .
Kashat's business rules significantly influence its algorithm’s operation for customers with no loan history by focusing on minimizing risks associated with their unestablished repayment practices. For these customers, have no prior borrowing data, Kashat resorts to cautious limits from initial scoring based on phone data. The models train on these types of data to estimate credibility, assigning lower loan equivalents to mitigate potential financial risks .
The PAR report plays a critical role in assessing Kashat's model performance by providing a detailed analysis of loans disbursement and repayment across various timeframes. Measures taken from this report include the number of loans disbursed per month, principal amounts, and the timeliness of repayments (e.g., 7 days, 15 days, 30 days late). These metrics help determine how effectively the model predicts loan recoverability and manage risks while informing strategic adjustments to the algorithm based on observed repayment trends .
The Kashat algorithm incorporates a score revision mechanism during a loan cycle that evaluates and adjusts user scores based on repayment behavior. Adjustments are made for events such as early, on-time, or late repayment of installments. For example, scores are revised if the first installment is repaid early or within a grace period, or if the second installment is overdue. These adjustments are reflected in the Kashat application interface, modifying the customer's credit profile based on their repayment consistency and compliance with set business rules .
The Kashat algorithm evaluates a loan applicant's eligibility and credit score by performing several steps: The customer requests a loan using the Kashat mobile application. The application then extracts information from the customer’s phone based on permissions and directs the customer to fill out a questionnaire. The user passes through a fraud rule engine, and if not flagged as fraudulent, the application calls the credit scoring API to assign a credit score. The scoring API returns the credit score and level/band, which determine the loan acceptance or rejection based on a set threshold for new and recurring customers. The algorithm also utilizes historical data where available to assign a credit score and limit amount .
For recurring customers, the Kashat algorithm utilizes around 60 features primarily focusing on historical loan repayment behavior. These include data points like payment history, past loan amounts, repayment patterns, and other personal loan-related metrics. These features are considered more robust than those used for new customers, as they offer a more accurate prediction of future repayment behaviors .
Kashat measures the health of its loan portfolio using the Portfolio-At-Risk (PAR) report, which evaluates loans based on their disbursement and repayment trajectories over specified intervals. Indicators include the number of loans disbursed, principal amounts, and repayment timelines (on-time, 7 days late, etc.). These metrics include the tracking of repayment performance, such as the proportion of loans and amounts repaid across different delayed intervals, offering insights into the portfolio's risk and recovery efficiency .