How can lenders best measure and manage credit risk, given the disruptive patterns in consumer behaviour over the last 18 months? How can large banks ensure their digital transformation programmes make them more competitive? What is the value of a diverse team and new ideas?
Last week a FICO team met with chief risk officers from some of the biggest UK banks to discuss these and other challenges, at our UK CRO Summit. We heard from thought leader Matthew Syed, author of Rebel Thinking, on how bringing together more diverse teams can unlock new approaches to problems and improve performance. Following this — and before an excellent dinner in a spectacular location atop Tower Bridge — we dug into topics suggested by the CROs present.
Managing Risk Models in a Crisis
One of the most difficult problems faced by risk leaders worldwide involves the changes in consumer risk, and how to measure them to make better decisions. Every major change in the economy raises the issue of risk model performance, given that models are based on risk patterns from a prior period. COVID-19 has exacerbated the difficulty of getting a true gauge on risk, given that consumers have received government support, such that their actual financial difficulties may be masked until that support is removed.
We know from past crises that robust risk models will continue to rank-order well even as circumstances change. But the actual risk level at each score band will change, making model monitoring and governance more critical. Some suggestions from our discussion, and from FICO’s analytics leaders, include:
- The frequency of odds-to-score tracking should increase for certain populations/products. Profiling analysis and tracking early performance can help identify future issues or sub-populations to focus on. Baselines, for comparison, should be taken from a period pre-COVID that reflects your best estimate of the future portfolio.
- Characteristic analysis that is conducted on a regular basis can be widened to include variables outside of the model. We suggest focusing on data that has the ability to show trends or change and that can be operationalised as early warning indicators or to aid decisions through overlays to existing models and/or strategies – temporary measures that demonstrate sound risk management practices. As an example, credit bureau delinquency variables in the model may be static but variables such as “number of enquiries” will show volatility regardless of support received.
- Affordability assessments can play an important part in understanding a customer’s resilience, what they can afford and where their need for credit of different varieties is. These assessments should become more commonplace throughout credit decisioning and not used solely at origination and for later-stage collections.
- Payment holidays should not impact score validity. The score will not reflect the temporary status but, if robust, will be reliable after the programs have ended. Determining the lag from the endpoint to when the model’s reliability should be back to normal will not be trivial, and will depend on the portfolio dynamics and lending appetite moving forward.
- Transactional data (PCA data or Open Banking) will be valuable – spend patterns and analysis requires a smaller timeframe.
- The criteria for the redevelopment of models remain the same – redevelopments should be prompted by the need to improve accuracy and/or stability because of significant shifts in population and/or characteristic distributions. Improved model accuracy, for COVID-19 in the near term, will be difficult to achieve through model redevelopment due to dependency on historical data. In future model developments, historical data that overlaps with COVID-19 should be avoided or used with caution since credit need/payment behaviour will not be representative. In addition, the redevelopment of models for particular use cases is not a priority, because of the time and effort required to get them validated and approved before they can operationalised.
- The immediate focus should be on revisiting decision strategies to make risk-targeted modifications and overlays, including changing dialler prioritisation rules, using more data at origination and increasing exposure selectively, with greater focus on a customers’ capability and desire to repay. For example, you might take those accounts that score just above a score cutoff threshold for approval – accounts that would normally be approved automatically – and review them using additional data, such as an income stability metric. Just remember to track and eventually remove that overlay at the appropriate time and when a consumers’ true financial stress is reflected in their data. This tracking should be added to the overall governance surrounding overlays and/or policy rules so that the ones used remain optimal and these short-term measures aren’t added to an ever-growing list.
- Decision models such as action-effect models, as used in strategy optimization, are valuable as you can efficiently perform scenario-based ‘what-if’ analysis, allowing you to understand the business impact of applying or removing overlays. These capabilities are part of FICO optimization tools and Business Outcome Simulator.
Gatherings like the UK CRO Summit give industry leaders a chance to tackle difficult issues like this together. I’d like to thank all the participants from last week, and I look forward to meeting with leaders this year in FICO’s programme of think tanks, round tables and other events — not least FICO’s global conference in May, FICO World. And of course, we are happy to discuss the best strategy for managing risk model performance and adjusting risk strategies.