Where Do Closures Cluster?
The histogram below shows the distribution of financial stress scores for operating institutions (blue) and those that have closed (rose). Closures cluster almost exclusively above a score of 85. The vast majority of the distribution is invisible to institutions in the "safe" range.
Financial stress score: enrollment trend (5-year CAGR) + tuition dependency (tuition revenue / total revenue) + days cash on hand. Scored 0–100, higher = more stressed.
How Many Are in the Danger Zone?
We grouped institutions into four tiers based on stress score. Over half of all institutions fall into the Critical tier, and that tier contains every single closure in the dataset.
The Critical tier is where all the action is. All 244 closures in the dataset occurred among institutions with a stress score above 75. The Low, Moderate, and High tiers have zero closures combined. This extreme concentration is what makes a simple threshold rule work so well.
Top 20 At-Risk Operating Institutions
These are the 20 currently-operating institutions with the highest financial stress scores. The pattern is clear: small enrollment, near-total tuition dependency, minimal cash reserves.
| Institution | State | Sector | Enrollment | Score | Tuition Dep. | Days Cash |
|---|
Source: College Scorecard and IPEDS 2015–2023 data. Stress score computed from enrollment trend (5-year CAGR), tuition dependency ratio, and days cash on hand.
How Good Is This Simple Rule?
The 3-variable stress score with a threshold of 75 captures 93% of actual closures (recall) with a precision of 33%. That means for every 3 institutions flagged as high risk, roughly 1 actually closed. The false-positive rate is high, but in an early-warning system, that's acceptable. You'd rather flag too many than miss the ones that close.
A simple model works here because the signal is strong and the dataset is small. A 3-variable rule with no machine learning, no tuning, and no black-box complexity catches 93% of closures. The ML model (Q5) improves this by ~7 percentage points in AUC, which is real but modest. The simple model is the right starting point because it's explainable and cheap to maintain.