What exactly Is AI-automated underwriting?
If you’re reading this post, you’ve probably heard of Zest AI and what we do—but do you understand it? There is a lot of buzz around artificial intelligence and machine learning, but there are also a lot of misconceptions. We’re here to demystify AI and show you how it can not only improve your business, but also give you the power to be more fair and inclusive in your decision making.
How does Zest use AI?
Zest models are more accurate than legacy scoring methods because we use more data and better math — but what exactly do those things mean?
More data is obvious. Instead of relying on a few dozen variables like traditional scores do, Zest uses hundreds of variables to paint a more accurate picture of a borrower. Just think of it this way: Those variables in the model are like the number of pixels in a digital photo — wouldn’t you rather have a photo that has several hundred pixels versus just a few dozen? A few dozen pixels might tell you whether the room is dark or light, but it won’t tell you much about what’s happening. Applied to a model, this additional resolution (the additional variables) can contain information that might make it more obvious whether a prospective borrower should be approved. Limit the number of variables and you just can’t “see” what’s really happening.
“Better math” requires a deeper explanation. Although many use the terms interchangeably, there is a difference between artificial intelligence (AI) and machine learning (ML). Machine learning is a specific field within the more diverse field of AI (think of them as a subsidiary and parent company, respectively). AI focuses on building intelligent machines and includes such things as understanding and generating language, robotics, planning, perception, reasoning, and acting in the world. Machine learning is the part of AI that focuses on learning patterns from data to make predictions.
Advances in machine learning over recent decades have been significant. It has delivered significant improvements in the quality of services like drive time estimates on your GPS, monitoring for fraudulent credit card transactions, improving the quality of Google search results, and making content recommendations, among many others.
However, it’s important to understand that machine learning is just the next generation of predictive analytics or statistics. Machine learning is more advanced, but it’s just learning how to make predictions from data, just like traditional statistical methods. What’s different is that the sophisticated methods developed by machine learning scientists enable vastly more accurate predictions than traditional methods.
Zest uses machine learning to build more accurate and inclusive credit scoring models. The “better math” used in these models not only considers a lot more information about a borrower (the “more data”), but it can also learn important relationships between each of the variables. Traditional scoring methods treat the (few) variables they use as independent, instead of looking at what they mean in combination. To continue our analogy, the old way is like looking at each pixel individually, instead of how those pixels make up a bigger picture.
For example, say a person has 99% utilization of their available credit. Under traditional models, this alone would be equated as high risk — but what is their total credit limit amount? Someone with high utilization and a $1,000 credit limit is a very different risk than someone with high utilization and a $100,000 credit limit. Zest’s machine learning models look at high utilization (or any other variable) in conjunction with other information available, instead of treating each one as independent.
How can AI-automated underwriting benefit your credit union?
As we know, using more data and better math will result in a more accurate prediction of risk. However, the data we use is tailored to the community and members you serve – as well as the ones you don’t. Our software is able to produce a Zest score that is more consistent and trustworthy than a traditional credit score – allowing you to confidently deploy capital to the members who need it most.
Auto-decision more borrowers
When you have trust and confidence in the score of your applicants, you can automate more — if not all — of those lending decisions.
Just think of some of the reasons why you might flag a loan for manual review — is it because, in some cases, traditional scores can’t be trusted? For example, you know credit scores below 700 are less accurate at predicting default risk, so you might choose to manually review those applications. Unfortunately, there are a lot of borrowers that have credit scores below 700 — Experian estimates as many as 40%, or approximately 120 million Americans, have scores below 700.
On the other hand, you can trust that the technology from Zest will provide an accurate assessment of the risk of a borrower each and every time. That trust has allowed some credit unions to increase their auto-decisioning rates by more than 90%. When you adopt automated underwriting, your members will be happier because they get a decision in an instant, and you will have freed up time for your underwriters to focus on more important things. Now your credit union can prioritize things like financial wellness and education programs to better serve the needs of your members.
Build more inclusive communities
Automating your decisioning process not only makes your underwriting process more efficient, but also more consistent. And using Zest software to power your lending decisions means every decision will be more fair and inclusive, because of the work that Zest does to spot and reduce bias.
At the foundation of every credit union’s ethos is equity and inclusion. Unfortunately, the truth is that all people have some sort of bias, no matter how well-intentioned. Even the data used to train models has bias because it’s based on biased lending practices of the past. Fortunately, that can be remedied by adopting AI-driven underwriting solutions.
Living up to the promise of inclusivity is the right thing to do, but it also makes good business sense. Just think of all the creditworthy people who may have been denied loans because those institutions were using legacy technology, but then come to your credit union and finally get approved. That creates a level of loyalty and gratitude that can’t be bought.
How do we know AI models work?
Remove bias from the equation
Legacy scoring methods are flawed. Unfortunately, for decades in this country it was common for financial institutions to deny loans to people based on a race, gender, sexual orientation, or any other of what is now considered a protected class. This is why it’s important to remove bias from a model before it’s built.
As mentioned earlier, machine learning also uses historical data to train models — so what makes Zest different from a legacy credit score?
We use a method called adversarial debiasing to remove bias and create outcomes that are both accurate and fair. The method allows the learning process to pursue two objectives: The first objective is to be as accurate as possible in predicting whether someone will default on their loan. The second objective is to be as fair as possible in the process. By making the learning process more aware of whether it is learning an unfair or biased pattern, we can render fairer decisions.
Resiliency, no matter the circumstances
One question that gets asked frequently is whether the AI model will change over time. The answer to that is no. Zest models make point-in-time predictions based on historical data, which means they aren’t learning from new data on their own.
Of course, now the question is, how do you know that the models will remain accurate during uncertain economic conditions if they aren’t learning from the most recent data? It’s because the patterns that indicate whether you’re likely to repay your loan don’t change that much over time. Zest models also make use of more variables, so they are more robust and stable to changes.
At the height of the COVID-19 pandemic, the U.S. government essentially put a pause on negative reporting to protect consumer credit scores. Now, the information normally used to assess risk and underwrite a loan — such as late payments — was no longer viable. As a result, credit scores increased to record highs even as the unemployment rate kept getting higher.
Now, you might expect that Zest’s software would also think these borrowers are less risky than they actually are — but machine learning models aren’t so easily fooled. During the pandemic, Zest models proved more accurate than credit scores. In a recent study, we looked back at the 2008 financial crisis and showed how, if at that time, lending decisions were based on a Zest score, the outcomes would have been more accurate than traditional credit scores, even during the worst economic cycle in recent history.
Conclusion
AI-driven underwriting isn’t some mysterious, all-knowing robot bent on taking over the world, but it is a powerful tool for financial institutions of all shapes and sizes. It saves you and your customers precious time, promotes a more inclusive economy, and can help you serve your community better even through the most turbulent times.
