Your early delinquency rates are ticking up. Your quarterly vintage reports show deterioration. The ground is shaking, but you don’t know where the earthquake is coming from.

Is it the economy? Your underwriting? That new marketing channel you launched? The honest answer: you probably can’t tell. And that could be costing you millions.

I’ve spent the past several weeks talking with some of the sharpest minds in credit risk, people who predicted the 2008 crisis two years early and navigated the pandemic without missing a beat. What they told me should make lenders rethink how they’re measuring portfolio performance.

The Blinders Problem

“What you’re seeing in your loan portfolio is not the full picture,” Kevin Moss told me. The former Wells Fargo and SoFi risk executive has a vivid way of explaining it: “It’s more like there are hidden messages. If you think about a horse that has blinders on, you’re not seeing the full landscape in front of you.”

Moss continued: “This is the way it’s always been. Horses have always had blinders on. And now we need to take those off and think differently about the problems we’re trying to solve.”

But here’s the uncomfortable truth: most lenders don’t even realize they’re wearing blinders. They’re tracking delinquencies, monitoring vintages, and reporting to their boards. Everything looks fine…until it doesn’t.

Joe Breeden, CEO of Deep Future Analytics, whose strategic analytics has been helping banks understand portfolio risk for decades, puts it more bluntly. When lenders come to him because their portfolio isn’t performing as expected, he finds they’re asking the wrong questions entirely.

“Most of them are tracking the right performance metrics, but they can’t identify underlying causes and they can’t do it early enough”, Breeden explained. “A young fintech might have completed their initial product testing and are ready to ramp up originations volume, but expanding inevitably brings adverse selection risk. The challenge is to measure this risk rapidly enough to adjust or change course so that you preserve yield.”

This is where things get interesting. Breeden’s approach doesn’t wait for 12 or 24 months to tell you if a channel is working. “We can tell you in a few months which of your channels are likely to perform better or worse than your existing book,” he said. “We’re measuring the net reality. Not what the economists or scores say.”

The Yield Optimization Gap

Here’s what really surprised me: very few lenders are actually optimizing for yield. They’re optimizing for loss minimization, which isn’t the same thing.

Pankaj Kulshreshtha, CEO of Scienaptic AI, confirmed this when we spoke. “Very few, there are two credit unions out of all our clients that have a very clear framework,” he said. “The lending officers have very clear framework for yield maximization and their targets. Typically though, what you see is most people are looking for loss minimization. They have a certain risk appetite and that is where they want to stay.”

The reason? “It’s complex,” Kulshreshtha explained. “The only prediction they currently have is probability of default models. But you need to have multiple models. You need to have models to predict the probability of activation. You need to have models to predict the likelihood of spend. You need to have models to predict the likelihood of prepayment. You need to have models to predict likelihood of revolving behavior [for credit cards].”

Brian Hughes, the former Chief Risk Officer at Discover and now a Director on Affirm’s board and an advisor to multiple fintech companies, sees this playing out across the industry. At sophisticated lenders, he explained, “You model the charge off curve. You model revenue. So you’re forecasting ROA over time on a monthly basis at a micro-segment level. And so at three months, you’re reading credit risk, you’re reading revenue, you’re reading ROA.”

But here’s the thing: “That’s a Capital One or Discover scale kind of operation,” Hughes noted. “Most lenders aren’t operating at that level of sophistication.”

Why Traditional Models Miss the Mark

The problem with traditional modeling approaches isn’t that they’re wrong, it’s that they’re incomplete. And that incompleteness creates blind spots that can cause problems.

Breeden illustrated this with an example from 2006, before the financial crisis. “The delinquency wasn’t bad, but the environment was really good,” he explained. His models spotted something others missed: losses were minimal, but they were way higher than they should have been given the economic conditions.

“What happened was the probability of default was still very high,” Kevin Moss explained, recalling his time taking over Wells Fargo’s home equity business in 2007. “But homes were being sold as the way for people to pay off their debt. So losses weren’t being realized. But when real estate values started to drop, people lost their way out.”

Breeden’s approach predicted this two years early because it separated credit quality from economic factors. “We’re not ignoring the economy, we’re ignoring the economists,” he clarified. “Because in the pandemic, what the economists know are GDP, unemployment. But the reality is forbearance programs, checks in the mail, all these things that they don’t track.”

His models worked perfectly through the pandemic for this exact reason, they measured what was actually happening in loan performance, not what economic forecasts predicted should happen.

The Cash Flow Advantage

One of the emerging solutions to this problem comes from an unexpected place: cash flow data (see my previous article on cash flow underwriting).

Hughes, who advises Nova Credit, explained how cash flow underwriting provides signals that traditional credit bureau data misses. “Cash flow gives you a different signal than the credit file,” he said. “It gives you a more timely signal as to when a customer is running into trouble and it tells you about their ability to pay. You’ll see trouble in the checking account well before you start to see trouble on a credit bureau report.”

This has enormous implications. “By the time someone’s reported delinquent on a credit bureau, it’s 30 to 60 days after they’ve missed a payment,” Hughes continued. “And before they’ve missed a payment, chances are they’ve lost their job or they’ve had an expense shock and that’s visible in the checking account.”

The practical application? “If you’re a credit card business and you see this happen, you certainly want to exclude them from any credit line increase campaign,” Hughes explained. “You also have the ability to reach out to help the customer before the delinquency shows up.”

Better yet, Hughes noted, cash flow data can help identify who’s actually revolving on their credit cards, one of the key revenue indicators. “You can look in the checking account data and you can see what they’re paying to their credit card, and then you can look at the credit bureau data and you can see what their balance is. You compare those two, you can reliably guess at who’s revolving.”

Learning Faster, Adjusting Quicker

The best lenders share a common trait: they learn incredibly fast.

Kulshreshtha emphasized this when discussing how Scienaptic’s approach differs from traditional modeling. “Any performance data that you are getting back, you are using to decide whether it is time to do a full blown retraining of the model or you can just live with few tweaks based on the profile changes that you are seeing,” he explained.

“If you have early delinquencies that are running ahead of the established vintage curve of losses, then you want to react faster,” Kulshreshtha continued. “That is the beauty of the new age machine learning, you can start responding to trends much faster than conventional scores would.”

Hughes pointed to buy-now-pay-later as an example of rapid learning done right. “One of the advantages of the pay-in-four product is that vintage performance can be read much earlier than that of a credit card, because the first payment is due 15 days after the purchase, and there’s no option to make a lower, minimum payment.” he said. “That is one of the innovations in buy-now-pay-later: a quicker read.”

For traditional lenders, Kulshreshtha offered stark advice for those experiencing portfolio deterioration: “Please make the change to your underwriting quickly. We can come in and help you and be commercially flexible. But you need to stop using the wrong toolkit ASAP.”

The Collections Insight

There’s another dimension to this that often gets overlooked: the relationship between underwriting and collections.

“Lending businesses are not so much about lending, as they are about collecting,” Kulshreshtha told me. “Once you have underwritten a personal loan, for example, an unsecured personal loan, money has just left you. If you do not do a good job underwriting upfront, then you will end up having accumulated losses over time.”

This connects directly to what Hughes described about pre-delinquency collections. “When it comes to collections, being the first to contact a customer usually creates payment preference for you,” he explained. “If you reach out early with a lower interest rate or a payment plan, you can be seen as helpful and have a good chance of beating out the other companies trying to collect.”

The challenge has always been that pre-delinquency campaigns were often NPV negative – you were wrong more often than right. Cash flow data and better modeling are changing that calculus.

What Makes the Difference

So what separates the winners from everyone else?

According to Kulshreshtha, it comes down to two things: “What is going into the funnel upfront is better and is continuing to get refined as changes in the environment come through. You got to learn very quickly. You got to see vintages, you got to slice and dice data in a nuanced way to figure out what is working with your strategy and what is not working and make those changes quickly.”

Moss agreed but added an important caveat about why change is so hard. “Most people in their business lives are struggling to just get what they know done,” he said. “There’s very little room to innovate. It could be budget, it could be time. It could be you gotta convince a whole bunch of other people – regulators, model governance people. And they all have veto power over change.”

But here’s the thing Moss emphasized: “You know, the only way you’re gonna win is if you have different approaches. Better product, different underwriting, different data sources, better use of the data. What’s your winning edge?”

Breeden, who has published most of his methodologies openly, expressed some frustration about adoption. “I don’t know any vendors in the US doing what we do, but there are plenty of banks who have taken the first steps internally,” he said. “The vintage modeling approach is used all around the world.”

When I asked why more lenders don’t adopt these approaches, Moss had a simple answer: “People are slow to acknowledge and accept change. They’ve done things that have worked for them, although they may be somewhat sub-optimal.”

Optimizing for Profit

In economic theory, a business will maximize profit up to the point when the marginal cost equals the marginal revenue. So for lenders, this means knowing exactly how much revenue each new loan will produce. 

But there is a problem here. As Moss puts it, “The problem as I see it is that it is a two- dimensional optimization (score and theoretical profit), where Breeden’s suggested approach is more multi-dimensional. The traditional approach is also backward looking, in the sense that you might be using an out of time validation sample, but it is 18-24 months old to align with the performance outcome of how the model was built.”

Breeden emphasizes the importance of not focusing on what was profitable before. “That is all backward-looking rather than asking what will be profitable with current economic conditions, current macroeconomic adverse selection, current account management policies, etc.”

When optimizing for profit you need to take into account prepayments as well as losses, and more importantly the timing of those.

Here is Breeden again. “Loss or prepayment early in the life of a loan is very expensive. Late losses or prepayments may make little difference. Most finance calculations I see just plug in annual rates, which is very self-deluding and misses differences and opportunities by segment.”

The Path Forward

The lenders who will be most successful going forward will be those willing to undergo a fundamental shift in how they think about portfolio performance.

“I think the people it resonates most with just have good portfolio intuition,” Breeden told me, explaining why his approach clicks with practitioners like Moss and Hughes. “I have a harder time with statisticians who’ve never heard of this kind of analysis and don’t accept it because it’s not what they learned in school. And they have no credit risk intuition.”

The key insight isn’t that your models are necessarily bad, it’s that they’re not asking the right questions. Are you measuring yield or just defaults? Can you separate credit quality from economic factors? Do you know which channels are producing your most profitable loans? Can you spot deterioration in 30 days instead of six months?

“It’s a different way to think about the same problem,” Moss said simply.

Maybe it’s time to take off the blinders and look at the full landscape. The ground might still be shaking, but at least you’ll know where the earthquake is coming from, and more importantly, what to do about it.

If you want to reach out to any of the experts quoted in this article, I will be happy to put you in touch. Just hit me up on LinkedIn here.