Leaning downhill: rethinking risk in lending

The counterintuitive path to risk reduction with AI-automated underwriting

Imagine you're skiing down a steep mountain trail. The wind whips around you as your speed increases and visibility diminishes. Your instinct is to lean back uphill, where it feels safe. Leaning downhill is frightening and seems fraught with risk. But this instinct — however natural — is misguided. 

In skiing, leaning backward increases the risk of falling, while leaning downhill enhances control and reduces risk, no matter how counterintuitive it feels. This analogy perfectly mirrors the lending industry.
‍

Hitting the slopes

In lending, the instinctive approach is to pile on policy after policy, rule after rule, on top of your model score. This might seem like a risk-averse strategy, reducing efficiency and automation in exchange for perceived risk reduction. However, this approach actually increases risk.

To understand why, we need to redefine risk. Traditionally, the lending industry equated risk with credit scores. A higher credit score meant a lower risk, and vice versa. But this perspective is flawed. It assumes that the industry credit score is an accurate and fair representation of risk, which is not always true — especially for middle to lower-tier borrowers. 

Risk should be defined as the likelihood of an applicant going delinquent and the expected loss given that delinquency. Through the use of AI, we can build powerful, accurate models to predict or accurately estimate this definition of risk for a population of borrowers. Of course, whether a given consumer actually defaults can’t be known with precision at the time of underwriting. 
‍

Bunny slopes vs. black diamonds

When an approvable applicant triggers a policy or rule, they're sent to manual review. This might make you feel in control of the risk, but the opposite is true. Instead of using your model, which has learned from millions of borrowers and hundreds of data points, you're relying on a human who can only consider 20 or 30 data points and who has learned from far fewer examples. Lenders need to leverage their models more to reduce risk and increase the number of model-based decisions.

This approach challenges the lending orthodoxy of the past decades. But just as we now trust our favorite map apps for directions instead of asking a friend, financial institutions need to rely more on their robust credit models. At Zest AI, we only make recommendations backed by data. Let's examine some in the chart below.

‍

This chart above shows that as the DTI threshold decreases (tightens) the risk actually increases! This goes against decades of lending intuition.

Relaxing certain policies and rules increases risk. This doesn't mean all policies and rules increase risk all the time. But if you have a powerful, accurate model, most policies and rules will increase risk most of the time. This is particularly true for rules providing redundant signals on data already considered in the model, such as the number of tradelines, length of credit history, and number of inquiry rules. Many rules are simply outdated, like minimum months employed, which is less relevant today due to increased job switching, especially among younger applicants.
‍

Acing advanced ski techniques with the right tools

Models have the advantage of being tested, validated, and monitored, while decision overlays often are not. As macroeconomic conditions change and models show deterioration, they can be refreshed with more recent data to maintain accuracy and effectively rank order risk. Rules and policies, however, often remain in place for decades, even when their creators have long left the institution. These rules and policies need continual evaluation to ensure optimization. At Zest AI, we work with our clients to evaluate which rules are necessary and which increase risk, despite the false sense of safety they provide.

It is time to lean downhill, embrace the power of AI, and redefine risk in lending. The future of lending lies not in outdated rules and policies, but in powerful, data-driven models that accurately predict risk and enhance efficiency.