Reducing loss rates for a global auto lender while automating credit underwriting
About the Client
Toyota Financial Services (IFL) is a wholly-owned subsidiary of Toyota Motor Corporation (TMC) Japan which delivers financial products, services or experiences through innovation, business transformation and new technology. The key target segments for TFSIN comprises of retail customers, customers who use the car for commercial purpose, small fleet and large fleet operators. Actify Data Labs has been the data science partner for Toyota Financial Services India Ltd. (TFSIN).
- Achieved 70% automated underwriting.
- Projected reduction in loss rates (60 DPD) by 1.5% – 3%.
- Retention model allowed TFSIN fine tuning dealer commission on second loan.
TFSIN faced a sharp spike in loss rates particularly in the ‘cab driver’ segment and was keen to automate the credit underwriting process and reduce loss rates. It also envisaged the creation of a new loan origination system that would incorporate the automated decisioning process.
For this purpose, Actify Data Labs conducted an initial consulting and scoping exercise to understand the change management processes that will be involved and the data that is available for developing a robust set of application scores.
Post the consulting exercise, Actify developed a set of four application scorecards:
- Commercial and small fleet customers, having credit bureau information
- Commercial and small fleet customers, not having credit bureau information
- Retail customers, having credit bureau information
- Retail customers, not having credit bureau information
An extensive data quality and data verification rules were used to identify erroneous historical data that were captured manually. It was also found that historical credit bureau data was stored in unstructured format. Actify used its unstructured data processing utilities to leverage that data to create credit characteristics from the bureau data. Multiple methodologies including statistical learning (logistic regression) and machine learning (Gradient bosting – CatBoost, Neural Network and Random Forest) were tested. Gradient boosting provided the highest predictive power; however, the same was not significantly different compared to logistic regression.
Hence, an easier-to-explain methodology like logistic regression was selected. Along with the probability of being bad (default), the model also provided top-3 reasons for the application being risky. These reason codes are used in the underwriting process for escalation and loan re-negotiation.
Post model development various swap set and scenario analysis were performed to determine the possible impact of the model on straight through processing, approval rate, loss rate and current underwriting processes. The impact of various policy rules was also analysed and those were used to create high and low side overrides. The results were discussed with the underwriting team, operations team, marketing team and the portfolio owner in an interactive workshop. Feedback from the workshop was used to create the policy rules and the modified underwriting process.
Using results from the model, it was possible to automate almost 70% of the underwriting process, while bad rates (60+ Days Past Due) were projected to go down by ~1%.
In addition, to the credit underwriting models and the credit underwriting framework, Actify also helped TFSIN, build an analytical framework for predicting retention of existing customers. The framework was aimed to proactively identifying current customers who will take up another auto loan within a given period of the current loan being closed. A retention data mart was developed in ADAPTify and a retention model was developed to predict customers who will take up a second loan. The top 10% of high probability retention score, was enough for identifying 60% of potential customers who would take up a second loan. This model helped TFSIN fine tune dealer commission on second loan; in addition, it helped them the internal target the right customers, thereby reducing effort and cost of targeting.
Workflow for building retention model
Right knowledge at the right hand.