In recent work, higher education researchers collaborated with the Federal Reserve to develop a predictive model that combines hundreds of institutional characteristics to estimate the likelihood a college might close. The model outperforms financial monitoring systems currently used by the federal government, offering a more nuanced understanding of financial distress in higher ed.
WHY COLLEGES CLOSE
College closures remain relatively uncommon, but financial distress is more widespread. Small institutions, particularly those that rely heavily on tuition revenue, have little margin for error when enrollment drops. Zack Mabel, director of research at the Georgetown University Center on Education and the Workforce (CEW) studying higher education finance, said colleges can experience financial distress for years without closure, and most institutions under stress ultimately do not close. This makes predicting which schools will shutter especially challenging.
“[I] think there are many different factors that contribute to why colleges close,” he said. “There’s no silver bullet.”
Phillip Levine, a professor of economics at Wellesley College, said closures should be understood as the most extreme outcome of broader financial pressures across higher education. For example, higher education institutions rely heavily on human infrastructure, which makes rising labor costs that outpace inflation put sustained stress on budgets. Over the past decade, inflation-adjusted tuition has been flat or declining, leaving institutions with persistent operating stress. When colleges can no longer cut costs without undermining their ability to function, closure becomes the remaining option.
“You can fire people, that reduces costs, but then someone’s got to stand in front of the room and teach the kids,” he said. “Then, the kids need food and support and other services. And at some point, it just becomes not sustainable.”
When closures do occur, they can be disruptive for students, employees and the surrounding community that relied on the institution to bolster the local economy.
CURRENT METHODS
Today, the federal government’s primary tool for assessing college financial health is the U.S. Department of Education’s Financial Responsibility Composite Scores (FRCS), a metric built from three accounting ratios that measure liquidity, equity and profitability. Institutions that fail the composite score or fall into a warning range may face heightened cash monitoring, which requires colleges to disburse financial aid to students before being reimbursed.
These measures function as compliance and risk control mechanisms rather than predictive tools. They are based on historical financial data and use a limited number of pre-specified indicators. They are also based on information from institutions that is not standardized, as not all schools use the same fiscal year.
The usefulness of these measures is constrained by a lack of timely data, Robert Kelchen, head of the Department of Educational Leadership and Policy Studies at the University of Tennessee, Knoxville, and contributor to the model, said. Federal data is based on audited financial reports and submissions to the Integrated Postsecondary Education Data System (IPEDS). This means financial stress becomes visible in federal systems only after it has been building for years. For example, he said, IPEDS data from the fiscal year 2024 only just became available.
MACHINE LEARNING
Machine learning models can address this problem by incorporating many variables simultaneously and allowing the data to determine how those variables interact. Rather than assuming which financial indicators matter most, these models test combinations of enrollment trends, revenue sources, institutional size, sector and other characteristics to estimate closure risk.
In the researchers’ model, enrollment declined and reliance on tuition revenue emerged as particularly strong predictors of closure. Other factors, such as institutional size, sector (public, private, nonprofit or for-profit) and even missing data also contributed to risk assessments. The model was tested against historical closure events and demonstrated higher predictive accuracy than FRCS-based monitoring alone — with 83 percent average accuracy compared to the federal methods’ 77 percent.
According to Mabel, the advantage of machine learning is not just that it processes more data, but that it does not require researchers to specify in advance how different factors should relate to one another.
“The model itself is going to figure out and test all the different ways in which multiple types of information can work together to provide a more comprehensive and more holistic portrait of the likelihood of closing,” he said.
