What About the Close Calls? In the Mortgage Market, the Behavior of One Group of Loan Applicants Is Particularly Enlightening

We can learn a lot about the home loan market by looking at the millions of mortgage applications submitted in the United States every year. Mortgage data, which are available thanks to the Home Mortgage Disclosure Act (HMDA), reveal many details about mortgage activity, resulting in a robust map of lending patterns nationwide.

Among the multitude of mortgage applicants represented in the data, is there a group of previously unidentified borrowers that might offer unique insights into the mortgage market? Three researchers - Hadi Elzayn, Simon Freyaldenhoven, and Minchul Shin - believe the answer to that question is yes.¹ The group they have in mind, known collectively as cross-applicants, comprises would-be borrowers who filed more than one mortgage application while seeking a home loan. In their paper, "Precision Without Labels: Detecting Cross-Applicants in Mortgage Data Using Unsupervised Learning," these researchers describe their novel method for identifying cross-applicants.

In the United States, roughly 22 percent of applicants apply for more than one mortgage.² These applicants are of interest for several reasons. For example, cross-applicants who had one application approved and a second (near-identical) application rejected can be thought of as "marginal applicants" on the lenders' decision boundary; identifying these marginal applicants may provide a better way to monitor current lending standards, enhancing our understanding of how lenders decide who to lend to. Cross-applicants can also shed light on the shopping behavior of mortgage applicants; looking at a given borrower's multiple applications reveals the factors being compared between lenders, thereby giving an idea of how much consumers stand to gain from comparing mortgage quotes.

However, it's difficult to isolate in the data those applicants who apply for more than one mortgage. To preserve the privacy of applicants, HMDA data do not include personal information. How can a researcher know which applications are from cross-applicants if there is no personally identifiable information associated with the applications?

The authors of this paper developed a novel method to overcome this challenge. In a nutshell, their algorithm first divides all mortgage applications into smaller groups (known as partitions) using variables that must be identical for two applications coming from the same applicant, such as zip code and gender. Next, it breaks down these partitions into clusters such that all applications within a given cluster are extremely similar (based on, for example, credit score, income, or application date). To evaluate and fine-tune their algorithm, the authors introduce to the literature a new method for evaluating unsupervised anonymous-record linkage, and they use this method to estimate that their algorithm achieves a precision of 92.3 percent - that is, 92.3 percent of all application clusters classified as cross-applicants indeed correspond to applications filed by the same applicant.

Analysts, policymakers, and business executives can all find practical uses for the authors' dataset. They might, as noted earlier, use it to study shopping behavior - with a potentially much larger sample size than traditional research methods such as surveys.

Marginal borrowers, given their position on the cusp of approval, also provide unique insights into the state of the mortgage market. These borrowers, being from the more vulnerable part of the applicant pool, offer a magnified view of the small inflection points that distinguish successful applicants from unsuccessful ones. Tracking their activity helps us monitor trends in lending standards and comparison shopping. In "Precision Without Labels: Detecting Cross-Applicants in Mortgage Data Using Unsupervised Learning," Elzayn, Freyaldenhoven, and Shin describe a new method for identifying these applicants, paving the way for future research into how the mortgage market behaves.