How AI can make lending objective, according to a tech leader

March 28, 2024 4:00 AM

In grade school, Sara Knochel already showed signs of turning into a future computer scientist thanks to a developing interest in math and languages.

In the years since, Knochel has applied those skills in an effort to help deliver solutions to some of the thorniest problems the mortgage industry is confronting today.

Her computer science degree led to an initial job at Loanperformance, which was later acquired by Corelogic, and eventually to her current leadership role at Candor Technology. But in between, Knochel earned an MBA, which drove her to other opportunities in financial services, where she saw how data analytics could be used in tracking loan performance across various segments.

The expertise initially brought her to Alpharetta, Georgia-based Candor as a consultant in 2018, and today, she serves as both chief operating officer and chief of staff.

Knochel spoke with National Mortgage News recently to discuss how the wealth of data the industry has today can be used to help mortgage lenders make accurate and bias-free decisions, while also addressing concerns the industry might have with trust in technology.

The interview has been edited for clarity and length.

One of your goals is to address the issues of bias in home lending and work on solutions. That's a tall task, but a very important one.

It could be like a lifetime of work. Yes, and it might end up being what Candor spends its entire existence doing. Our core product, what it does is it ensures that a loan and the document or data underlying it adhere to investor guidelines. And that's it in a nutshell.

If you can do that, what you can ensure is that all similarly situated loans are always going to be treated exactly the same if you have a machine doing it and not doing it with any pattern-based learning. But doing it with an objective approach.

Our AI is something called, "Expert System." It's not a trained dataset kind of technology at all. But what it does do is it gives you consistency and definitiveness. You know if something is meeting guidelines or not.

Candor is just here to check the guidelines. The lender is the one that gets to make the decision.

Explaining that to someone who maybe knows nothing about AI, how does it do that?

We might literally have to program into it investor guidelines, and we have to ingest all of the data that a human underwriter will look at — so the loan application, the credit report, the AUS [automated underwriting system] — and then all the data off all the supporting documents.

Once you get all that stuff in, it's like triangulation. Do all the documents match the application? And is what's on the application abiding by the investor guidelines? You've got the right amount of income, the right number of years, the right amount of assets — that's kind of it. The data ingestion is a big part. We do have to get data off documents.

But really, we do something called cognitive mapping, which is how we create the Expert System, and we actually have a team of underwriters, but they don't touch loans. They write the logic for the system. They basically create the brain. And they do this activity that we call cognitive mapping, where they map out how you think through checking a particular guideline. What's the path you've got to go down? What is all the data that you have to know? What are the things about the guideline that you have to know? How do you check it all and test it all to make sure that it's correct, and then what do you do if it's not correct?

What is the right remedy that you have to suggest? They map all that out and then our development team implements it, and they work very closely together.

Cognitive mapping is actually a psychology term. I think what we realized is the activity the human underwriter is going through is very detailed and complicated, but there is a set of definitive rules that guide it. That really makes it the right kind of problem for expert systems to be applied to.

Not just for this current job, but In general, what drove you towards technology?

I had this math teacher in the sixth grade who said he thought that I might like computer science because I liked math. And that I might like languages going into high school.

And then as I was trying to figure out what I wanted to do when I went to college — what I wanted to specialize in —I kind of knew computer science was one of those fields where you would learn hard skills you could use that you could pretty much always be employable with. But at the same time, it's applicable to so many different business types of problems. The world's kind of your oyster. You can take this skill set, and you can kind of end up doing anything — a nice mix of being practical, but also exactly what you might want to be interesting.

I really got fascinated with data, loved learning about the structures of data, the relationships of data, what you can do with them. Continuing down the mathematics path, I did a lot of things that I guess you would call artificial intelligence. I don't know if they were calling it that just yet.

But I did learn some machine-learning kinds of things. But also, I learned that AI was really a very big space. There were a lot of different kinds of tools in that world, and the machine learning and the training-based ones that are getting a lot of notoriety right now — it's only a small part of the world for AI.

There's just so many different kinds of things in that world. Using data to train a machine to recognize a pattern is a very powerful tool. But there's other things that qualify as AI that don't actually do that. It just kind of depends on what problem you are trying to solve and what tools are available.

Would you call all the analysis — all the data that you look at and your underwriters are feeding — is that considered artificial intelligence? It can be a wide term.

It's still a kind of artificial intelligence. What's interesting though is if you look at the CFPB [Consumer Financial Protection Bureau] definition of AI today, what they've defined as AI is machine learning. So by the CFPB's definition, they actually wouldn't consider it AI. But if I think about what the academic definition of AI is that I learned in school — technically we're AI.

I think that's a really important question to ask. It's just to understand again — to get down to plain language, what exactly are we talking about? We have a model that takes data and makes an estimation, or are we talking about a rules-driven system.

The mortgage industry has been a little slow on technology adoption in general for a variety of reasons. It could be budgetary concerns. It could be just a wait-and-see attitude, but how do you think the industry or companies like yours can build trust in AI?

That's a great question. One of the reasons we provide [a] warranty is we want to build that trust. I think that has helped a lot.

I think a lot of lenders have been given promises about technology, and when it doesn't pan out, they're the ones that are out of luck. There's really no one to hold accountable.

As much as you can measure ROI [return on investment], I think that's important. Lenders have been a little slow to adopt new technology, but I don't think it's because they haven't necessarily wanted to.

I think what they're finding is, when they have added a lot of technology, they haven't gotten a less expensive process. It's costing more, not saving them. So they're really struggling with scrutinizing the ROI out of each piece that they add to their process. That has to make sense.

People do have concerns about regulators and what regulators think. I think they want to know what regulators say about certain things. I think that lenders that are more savvy about the differences between technologies feel more comfortable moving ahead.

Did mortgage and housing find you versus you seeking it out?

I found it, but then I left and it came back. It's funny, one of my jobs when I was in college getting my computer science degree was — I found a posting to be a junior loan officer at a brokerage, and this was during the subprime credit crisis. That was an exciting time to be a junior loan officer. So I had that summer job. That was my only taste of the industry to start.

I got my computer science degree. And I did end up going into a company that did mortgage. I went to Loanperformance.

Residential mortgage — whether it's secondary market data, what happens after, or now being in fulfillment operations — it's pretty much always been that.

There was a brief period of time when I decided to get an MBA and learn more about the business problems that we use technology to solve, and then I ended up consulting for a while.

But again, that period of time, it was all data and analytics focused, and I saw a lot of interesting problems that helped me when I got to Candor in terms of how we just make our own product, and also the kinds of problems we try to solve for clients.

And so, I look at that time that I left mortgages and saw some other things as having been really important to my development, able to come back with a more well rounded skill set.

For most of my career up until Candor, I knew a lot about what happened to loans after they were made — modeling their performance and things like that. But I really got interested as to what happens in fulfillment and origination over the last few years being with Candor.

For tech enthusiasts like you, what paths in fintech are there that you found helpful?

The interesting thing about technology is that it's always changing, and if you're in it, you're always going to be learning. While you can get foundational degrees and things like that, I think what's really important is going back to "Are you understanding the problem that you're trying to solve? And then do you know how to evaluate what the right kind of tools are to solve that problem?" To me, that is key to successful technology.

Where does the data Candor uses come from?

The good thing is we don't actually use any data to make the system — it's not trained on any data.

But when a loan goes through, it's all the data that comes with the loan, so the application, the credit report, bank statements, tax returns, that kind of stuff. And it'll go through the system, and then it'll make sure all those things match up to what's on the loan application. Then it'll check investor guidelines.

Another really important thing we realized is that your answer is only as good as the information you put through the process. So you have to have a really high degree of confidence in the data coming in.

How much available data is out there?

I will say our system generates a lot of data every time it touches a loan. Every time it touches a loan, it acts like it's never seen it before and it reassesses everything.

And every time it runs, we take a snapshot of all of that, so we know every data point that went into the system. We also know all of the results that came out of it, and we take that snapshot every time it touches a loan. I think to date we've had over 500,000 loans, but we probably have 2.5 million snapshots of these loans it touched going through the system.

And I would say probably every time a loan goes through our system, we probably save 10,000 data points about what was on the loan, and also ask what did our systems say about it? What conditions did it create? What calculations did it make? It starts generating a lot of data.

If anything has changed, we always have a picture of exactly when. That becomes important because we warranty whatever it touches. And if a client ever says "Hey, I've got an investor who wants to put back this loan," we go and we pull that history.