Impact

• Improved accuracy of decisions led to reduced delinquency & improvement in portfolio performance

• Established foundation for Reject Inferencing & Cross Selling Opportunities

Challenge

A premier APAC-based NBFC’s pre-existing incumbent process for credit risk assessment was driven by a rule-based algorithm that was posing problems in maintaining consistency and timeliness. The field verification data used for the process proved inadequate and there was heavy reliance on bureau data to make up for the inconsistencies. The dependency proved problematic as the bureau data did not comprise all the required information about the target population.

Approach

To enable operational efficiency, TheMathCompany collaborated with the IT & digital team of the organization and developed a customized credit risk assessment application that gave an all-encompassing view of customer data, generated a score card to gauge suitability of credit applications and offered real-time insights. Critical lines of business were identified for implementation, data quality was improved to enable efficiency, and data sources were strengthened and diversified. The ML-driven solutioning would help improve modelling accuracy and stability, and the time taken from information capture to response.

Client Testimonial

"The flexible solution designed by the team enables routing all applications through the scoring framework saving a significant amount of time and money. This process has enabled us to become future ready. "

- Head of Analytics, Leading NBFC, APAC region

Solution

A customized application was built for the client to enable improved data capture, track pertinent factors that determined credit risk application acceptance/rejection and better train models to improve scorecard accuracy.

To facilitate the same, TheMathCompany re-designed the existing survey questionnaires being used for field verification visits by taking into account quantifiable risk metrics. This way, all the applications and visits would be documented.

The team also worked towards setting up a customized model design that accounted for business requirements and market parameters, and established guidelines to facilitate effective model monitoring, specifically focusing on input data quality controls, business strategy changes and model health KPIs.

Here's an overview of process improvement measures introduced by our TheMathCompany solutioning team:

                                                                                  (Fig i): Overview of the process improvement

Efficiency Enablement: Out of the 8 business lines, three critical lines of business were identified, and credit application scorecards were developed for the same. The dynamic nature of the business was taken into account, and observation and performance windows were chosen, accordingly. Strong data sources including those relating to application, bureau and field visits were identified to reduce dependency on incomplete data.


Optimum Deployment: A variety of math-based techniques such as binning, group-wise segmentation, and classification models like logistic regression and decision trees were developed in the process of scorecard generation.


ML-driven solutioning: By leveraging ML techniques, TheMathCompany further diversified application in application scorecards, such as boosting accuracy through surrogate models, improving model stability with ML based sampling & k-fold validation techniques.
The resulting application scorecard had a 93% model prediction accuracy, which would only improve with time. The results generated through the scorecard helped decision makers to extensively assess rejection and improve trust in credit quality. Furthermore, the time taken from information capture to response was delivered within 120 seconds. The existing default rate was 15.82% and this was brought down to an estimated 8.8%, i.e., brought down by almost 50%.

• Applications scored with the model to determine default probability & assigned a score 

• Based on the model scores the top 20% (highest chance of default) captured 55% of total defaults

• The model is expected to improve the default rate to 8.8% (about 50% improvement by sacrificing 20% sourcing volume)

share case study

facebook linkdin twitter

Want critical data insights to unlock business value?

Get Started
mobilelogo
Linkedin Instagram Facebook Twitter
Cookies policy Privacy policy