In 2020, the Credit Awards judges voted for Nationwide Building Society to win the Credit Modelling & Risk Team of the Year. Details from the entry are included here.
There is an opportunity for banks and building societies to find new ways to use the wealth of data they collect across their business in credit decisioning, to enhance the lending process.
Through 2019 and 2020, Nationwide Building Society’s Retail Decision Modelling team did just that, developing several pioneering initiatives that resulted in significant improvement in credit decisioning across retail products.
The team’s ground-breaking approach drew on advanced analytical solutions to help uncover innovative ways of generating value for the business. The team identified that transactional data from mortgage and current account customers could be integrated into key business decisions.
The objectives were to:
• Roll out industry-leading machine learning into wider products;
• Exploit transactional data;
• Identify big wins for the business by challenging existing processes.
The team’s pioneering approach was structured to deliver two main achievements:
1 Delivery of machine learning in credit scoring
In 2019, the decision modelling team delivered the first machine learning credit score for a major financial institution in the UK or US that could be used in live decisions. The use of advanced analytical models played a part in the development, but the team also worked closely with the risk and systems team to create a solution that could be implemented in legacy systems and fit seamlessly with underwriting processes.
The pilot model was expected to reduce losses by up to 10%. However, the team went on to deliver a second machine learning credit scoring model which was forecast to reduce losses by 15% while maintaining volumes. This improvement in the reduction of losses was a result of the integration of machine learning into the ‘reject inference’ aspect of the model build.
2 Integration of transactional data into business decisions
This second element involved pioneering work on optimising decisions through the better use of untapped transactional data in the business.
Notable achievements included replacing limited Office for National Statistics data for household expenditure with a new approach that leveraged current account transactional data without requiring time-consuming and complex transaction categorisation. The result was improved affordability calculations across mortgages and unsecured lending.
The team also seized the opportunity to tap into mortgage transactional data to support the targeting of customers in financial distress for collection purposes. Reviewing the performance of this process after implementation, the team saw demonstrable reductions in provisions due to a practical idea, and positive feedback from customers to the team’s colleagues in collections.
Nationwide’s decision modelling successes
In essence, the entire strategy in credit modelling and its outcomes could be summarised as:
• Using mortgage transactional data to drive improved performance in collections, both in provision reduction and customer experience;
• Employing current account transactional data to assess household expenditure in affordability calculations for all secured and unsecured products;
• Working with machine learning to build a first-of-its-kind customer score;
• Recommending that one million customers become eligible for pre-assessment on the building society’s unsecured products.
Nationwide’s decision modelling team are continuing to work on the third and fourth builds of a machine-learning credit scoring model, including a proof of concept in applying into secured lending credit scoring.
In addition, the team has run ‘hackathons’ for the whole risk department at Nationwide, using the power of the crowd to tackle business problems with data.
Increased demand for modelling has also spurred the team to extend advanced analytics into money-laundering targeting, which has identified improvements in targeting suspicious behaviour through advanced analytics.