Recently, Paul W. Harris, General Manager, Mortgage Analytics, appeared on PRGORESS in Lending’s Lending Buzz podcast to discuss some of the nuances in automating mortgage fraud detection and fraud for housing. Listen to the full podcast here or check out the conversation between Paul and Tony Garritano, below.
Q. Let's start off with a general question. What would you say is new in mortgage fraud technology?
A. Well, there's a number of common words that you hear: AI, machine learning, understanding risks earlier in transactions via APIs and then lending channel configuration. We've developed new models that leverage machine learning, pattern matching and AI to better detect new mortgage fraud risk, like synthetic identity. With component solutions via APIs, we're able to solve for a single variable like identity, or REO or industry watchlist checks. Those can be integrated into point of sale systems or other parts of the mortgage process to improve pull through or address targeted risk. Then, finally, we've customized versions of traditional fraud tools by lending channel so we can help correspondent TPO lenders, retail, and home equity.
Q. One common gripe with fraud technology that I hear a lot is that it generates false positives. What are false positives, and what are some common examples?
A. Well, let me start by saying legacy fraud tools do work, and they are one of the main reasons why fraud is so low today and has been for the better part of the past decade. But having said that, legacy solutions are designed to detect more than fraud. They also look for compliance risk and data integrity issues. Naturally, when you do that, you tune the models to be overly sensitive to certain variables and can generate a higher number of false positives.
To answer your question about false positives, though, it really is a variance, which incorrectly indicates that a particular condition or attribute is present. Think about common name matching. A name like Smith or Jones matched on an industry watch list or an eligible list can be considered a false-positive once you match it up to your borrower. Another example is REO. It can create unnecessary review of loans that, in reality, require a lower level of diligence related to risk, which can slow the approval process and create trust and adoption issues with the users.
But we found a better way. One of those mentioned earlier, customization, a regular calibration of a traditional fraud tool, and/or new tools like our AI Score. After testing millions of loans with alerts and actual frauds, we know which alerts are likely to generate a false-positive result. We also know which alerts are highly predictive of both fraud, as well as EPDs in the riskier score bands. Lenders can then focus their efforts on the loans that are most likely to be fraudulent and have high risk of EPD and minimize the review rates. The AI Score was really designed with larger lenders in mind. They're making intelligent acquisitions and dealing with ever-increasing volumes. The key benefits are obviously cost savings and efficiency, but ultimately decreased risk because the AI Score runs some models around things like synthetic ID and employment.
Q. When we talk about fraud, mortgage fraud is at historically low levels right now. How do we keep it that way going forward, in your opinion?
A. Well, mortgage fraud is at historically low levels. It's really been thanks to a heavy refi market, conservative loan products, thorough underwriting practices, and, one of my personal favorites, leveraging best-in-class fraud tools. The best way to prevent fraud is really with a solid defense. That's a combination of human skill and technology. On the human side, it's awareness, diligence and training. On the technology side, it's up-to-date prevention tools. But we need to avoid complacency, especially in this current high-volume origination environment.
While we're not seeing the levels of fraud we saw during the great financial crisis, there was actually an increase due to the pandemic, which has a lot to do with the increase in remote operations and digital mortgage activity. In 2020, a vast majority of mortgage transactions weren't fraudulent, to your point, but we did see the number of attempts to defraud mortgage lenders rising from 2019, as well as the number of successful attempts.
The industry has also seen a sharp rise in loan manufacturing defects, according to the most recent ACES' Mortgage QC Trends Report. These are primarily the result of high volumes, changing guidelines, and less seasoned staff. So we know fraud is down, but defects are increasing. As we shift to more of a purchase market, we anticipate a rise in that fraud due to the large number of participants in the transaction. We also know from industry data that, on average, for every dollar spent on fraud and compliance, lenders are protecting themselves against $3.50 in losses. That's up nearly 8% from 2019. So, the current practices are working.
Q. Well that's good news. Last question, before I let you go, Paul. Talking about current events again, are there new challenges we're seeing in verification that come from our current work-from-home virtual environment that you're seeing?
A. Absolutely, and I would put these new challenges in two different categories. One is COVID-specific. From the pandemic, you've got identity verification. Synthetic ID has become more of a concern for lenders, as well as employment misrepresentation. Think about hospitality, restaurant workers and independent contractors in the gig economy with really hard to verify employment. As borrowers exit forbearance, you're going to see lenders looking at REO disposition, which, as we've seen in the past, leads to occupancy fraud and sometimes not arms-length transactions.
Then, on the technology front, from consumer banking to mortgage, we've observed things like malicious bots getting more aggressive with malware cyberattacks. These are things that are happening in consumer banking and will end up in mortgage. We know that fraud evolves on a real-time basis and we anticipate that there will be additional unknown challenges that we have not yet observed.