Skip to content
Lab · ADM Case Study · Education

Which Colleges Will Close Next?

Over 200 U.S. colleges have closed since 2015. The next decade will be worse — and the closures are already visible in birth certificates from 2005.

We analyzed real College Scorecard and IPEDS data across thousands of degree-granting institutions with three tiers of closure prediction models. A simple 3-variable stress score catches most historical closures. Gradient boosting adds modest value. For monitoring decisions, the simple rule is the one you'd deploy.

Same data. Three models. The simplest one is the most defensible.

6,000+
Institutions
200+
Verified Closures
-15%
18-Year-Olds by 2029
39%
Say College Not Worth It
The ADM Thesis

Three Models, One Lesson

Each tier adds complexity. The question is whether that complexity adds value — or just noise.

Tier 1: Screening Rule

Three financial ratios: enrollment trend, tuition dependency, cash reserves. Catches the majority of historical closures. Fast, interpretable, deployable today. For monitoring, this is the one to use.

Q1: Risk Score

Tier 2: Multi-Variable Model

Adds demographics, geography, completion rates, default rates. The geographic signal — regional birth rate decline — explains more variance than school performance. Valuable for targeting interventions.

Q2: Geography

Tier 3: Gradient Boosting

Gradient boosting on 17 features. Real but modest improvement over the rule. The interaction between online share and sector is a genuine insight the simple rule misses. Worth it for deep-dive triage, not for screening.

Q5: ML vs. Rules

For a state higher ed board monitoring its institutions, the 3-variable stress score is the most defensible model. It's interpretable, it doesn't overfit, and it uses publicly available data. The gradient boosting model does better on paper, but with a small closure training set, the confidence intervals are wide. More complexity didn't clearly earn its cost here.

Honest Assessment

What We Don't Know

Small closure sample. Models trained on ~200 closures over a decade. This is a small-sample problem. Confidence intervals on recall are wide, and the gradient boosting model's advantage over the simple rule may not be statistically significant.

Financial data limitations. Two key stress score inputs — tuition dependency and days of cash on hand — are estimated from IPEDS Finance surveys where institution-level data wasn't available. Where data was missing, we used sector-level averages. This is declared honestly as a fidelity choice throughout the analysis.

Enrollment trends limited to recent years. Historical enrollment data covers a 5-year window. Institutions with longer decline trajectories may look stable in this window.