Implemented a centralized fraud detection system for retail loan applications at the time of screening.
Impact of the Engagement
The predictive accuracy of the deployed model observed an improvement of hit-rates by 50% compared to the earlier rule-based system.
ClientThe client is a leading provider of financial services across consumer and wholesale businesses. It has a comprehensive product suite to meet multiple financial needs which includes Consumer Lending as well as Corporate Lending.
High Non-Performing Asset (NPA)The retail asset portfolio of the client had a high NPA % which was largely contributed by fraudulent customers providing incorrect data at the time of applying for credit.
Challenge and Approach – Manual VerificationThe Risk Intelligence and Control division of the client was facing huge challenges in controlling fraudulent activities. In addition to this problem, the manual screening of the applications made it impossible to authenticate them. Consequently, the existing model saw deterioration in its ability to address the evolving complex nature and variety of frauds. Affine carried out a thorough analysis of the financial institution’s existing fraud management processes, data assets and historic fraud cases to create an amalgamated information to predict fraud:
- Data was collated from various internal and 3rd party sources to create an analytical data set for modeling and scoring.
- Primary focus was given to form the definition of the fraud events which were aligned to the financial institution’s existing fraud management processes.
- In addition, emphasis was given to strike a balance between mitigating risk, and adding friction that causes customer abandonment.
- KYC process was ensured to know enough about the customers to flag suspicious activities, and if necessary, have enough information to support an investigation.
- A range of advanced analytical and machine learning techniques were applied to compare fraud detection performance using historical data.
- An ensemble model combining three shortlisted models was used to score new loan applications for fraud probability. The alert mechanism was incorporated for the investigation team to take further actions
- The model was also designed to incorporate tweaks based on regulatory guidelines.
Outcome – Improved Hit Rate
- A centralized fraud detection system was implemented for retail loan applications. The usage of model was rolled out to all the outlets across the country.
- The model scoring codes were integrated into the Loan Origination System being used by the client. The run time processing of new applicants included generation of fraud scores using Affine’ s algorithm.
- Accuracy of more than 80% true positive cases enabled the client to reduce manual scrutiny of applications and thus improving the error rate.
- By identifying the applications that are more likely to be fraudulent, the model helped the client to minimize losses.
- Fraud prevention ensured faster, precise, and better yielding decisions along with reduction in processing delays.