Author | Vinay Bhaskar
Traditionally risk managers have responded to recession predictions by simply reducing risk appetite. This is roughly how that works - once there is enough buy in to the recessionary scenarios internally, risk score cut-offs are increased, and underwriting strategies are tightened. This results in an immediate drop in approval rates in the hope that when the real slow down hits, the losses will not be too bad. We believe a better framework to make lending businesses recession resistant and sustainably profitable is now available. There are multiple dimensions to this framework, but in this post, I focus on opportunities related to credit scoring.
Credit scoring methods have seen several generations of evolution, starting from intuition-based weightages to regression-based techniques. Machine Learning (ML) based techniques have been tried seriously in the last few years.
As per a McKinsey survey, respondents using AI in risk indicated they are seeing the great value. More than half of respondents report significant value from using AI in risk processes, and while the adoption of AI is still in its early days, the results suggest that it’s already reaping meaningful rewards.
The time is now ripe to make a definitive leap to industrialize next generation of credit decisioning using AI/ML. I believe the following are the important features of Underwriting in the years to come:
1. Advanced ML/AI techniques become the new normal: The benefits ML based predictive models are now broadly recognized. We have been able to find upwards of 10% incremental approval rates in all large retail lending portfolios (cards, unsecured loans) we work with. This is substantial impact by any standards. Even a few experts have genuine concerns about explainability and overfitting by ML based PD models. Fortunately, though, consensus is quickly emerging that most concerns with usage of ML / AI can be empirically addressed. I believe if done right, keeping in check the evangelical zeal, experts and business folks will get the comfort that comes with positive experience and more importantly with realized business impact. And lest we forget the legal and compliance comfort that is addressed with respect to AI in credit underwriting by the GAO (Government Accountability Office)who issued two reports (in March 2018 and December 2018) promoting or recommending interagency coordination on flexible regulatory standards for nascent financial technology (“Fintech”) business models (including through “regulatory sandboxes”) and the use of alternative data in underwriting processes.
2. It is a sin to have good models on shelf: I am surprised how often decision science teams at banks do fantastic work to build an advanced model, but it doesn't get implemented for the lack of internal buy in to ML/AI. Using a sharper decisioning instrument as soon as it is ready should be quite obvious; Google, Amazon, and many tech firms update their already sophisticated algorithms daily. Usage of sharper models creates immediate impact on growth and customer experience and on losses in due course; we need to create excitement about it in the organizations. There is an opportunity to reduce the model build-to-implement cycle time to a few weeks from the months it currently takes. Keeping glistening sharp knives in a locked cupboard isn't good for kitchen performance! Despite the nascent stage of AI adoption, its benefits are already being realized at many large banks across the globe. According to Deloitte research, “banking and FS executives found that investment in AI helped them reduce production costs by 13%. Additionally, executives reported a 17% average revenue increase in the area of their AI initiatives.
Friction in implementing ML/AI algorithms must reduce: There is every evidence that the algorithms to predict customer behaviors will continue to evolve. In order for underwriting instrumentation to stay sharp, experimentation with new algorithms is a must. Implementations systems, currently in place, have become too complex and clunky.
As per Bottomline technologies survey of the decision makers at AFP 2019 conference, nearly half of respondents (45%) noted that integration into existing technology is their biggest challenge when it comes to greater adoption of AI in banking.
Technology to allow this experimentation is now accessible, it just needs to be brought in regular practice. Reducing the implementation friction will enable faster time to market for advanced underwriting strategies.
4. Models must update more regularly Customer expectations and behaviors are changing rather rapidly. It is only natural that the predictive models be revised more dynamically. Data science talent availability has been a constraint, that is now being relaxed because of new tools and open source software. Arguably credit modeling requires a level of expertise and domain knowledge, but methods to enable more frequent model updates are now available.
These 4 elements of Underwriting, I believe, are important to the evolution of credit decisioning. There are reasonable concerns around each of these areas that would need to be discussed and solved for. Like any generational change, this will take dialogue, engagement and shared vision. We believe timing is right to elevate this discussion, especially as we are heading into slower growth.
Citing the fantastic analysis of AI and ML technologies in credit by Simeon Kostadinov, the banks and financial institutions have to address four additional critical challenges with respect to the technology implementation itself:
Biased data — all supervised ML algorithms assume that the future would follow similar patterns as the past. Therefore, a rich data strategy is key to extensive AI/ML adoption.
Interpretability — machine learning algorithms are considered “black-box” due to the inability to understand the reasoning behind individual decisions.
Regulations — in addition to ECOA, Congress has passed “The Fair Credit Reporting Act (FCRA)” which aims to maintain fairness in credit reporting and ensure consumer’s privacy by protecting or using certain information
Scalability — once a model is proven to perform well and does not interfere with any regulatory requirements, the challenge of scaling it up needs to be overcome
There are reasonable concerns around each of these areas that would need to be
discussed and solved for. Careful engagement with regulatory issues raised by new technology and practices across a range of requirements and contexts will be important to the development and expansion of sustainable credit programs built around AI and big data. Like any generational change, this will take dialogue, engagement and shared vision. We believe timing is right to elevate this discussion, especially as we are heading into slower growth.