What Can Public Data Say About College Closure Risk?
This study began with a synthetic screening cohort, then tested the same ideas against public College Scorecard and IPEDS records. The two cohorts tell very different stories.
The real-data output includes 1,516 institutions but only 6 observed closures. That is enough to expose weaknesses in the original models, but not enough to rank institutions confidently or support a deployment decision.
Same question. Two cohorts. Very different confidence.
By Michael Key · ORCID
Six Questions, Three Models
Read the Full Study →
Six investigations covering financial stress, geographic demographics, survival thresholds, the online pivot, whether gradient boosting beats a simple rule, and how well our initial synthetic assumptions matched reality. Plus the fidelity ladder that makes the ADM thesis concrete.
Explore the Data
Look up any institution, step through the fidelity ladder, or calculate your degree's ROI.
Institution Risk Screener
Type any college name. See its stress score, enrollment trend, demographic headwind, and exploratory model scores across all three tiers.
Fidelity Reveal
Same question at three fidelity levels. Watch the answer change as we add demographics, competition, and gradient boosting — and see where the extra complexity stops helping.
Degree Value Calculator
Pick a school type and major. See the payback period, lifetime earnings premium, and AI automation exposure. Earnings calibrated to published College Scorecard statistics.
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. In the real-data output, the reported threshold detected 0 of 6 observed closures. It remains interpretable, but this result does not validate it.
Tier 2: Multi-Variable Model
The logistic model reached 40% recall and 1.6% precision in cross-validation. With only six positive cases, those estimates are unstable.
Tier 3: Gradient Boosting
Gradient boosting reported an AUC of 0.901, but 0% recall at the selected threshold. The ranking signal is exploratory; it is not a calibrated closure probability.
No model in the current public output is ready for operational use. The useful ADM lesson is the correction itself: promising synthetic results did not survive contact with a sparse real-data cohort. The next fidelity step is better outcome coverage and external validation, not a more elaborate model.
Six Questions
Which Colleges Are Most at Risk?
The stress score is easy to inspect, but the selected threshold missed all six observed closures in the real-data output. Interpretability is not the same as validation.
How Much Does Geography Matter?
The six observed closures are spread across four regions. That sample is too sparse to support a geographic prediction claim.
When Does Decline Become Fatal?
Small enrollment remains a plausible screening signal, but only one of 93 institutions at the reported two-part threshold was observed closed.
Did Online Colleges Win or Lose?
The real-data cohort contains too few closures to estimate an online-by-sector interaction reliably. This remains a hypothesis, not a finding.
Can Gradient Boosting Beat the Stress Score?
Logistic regression detected some held-out closures; the stress rule and gradient boosting detected none at their selected thresholds. Six positive cases are not enough for a deployment comparison.
How Well Did Our Assumptions Match Reality?
We built this study first on synthetic data, then re-ran it on real College Scorecard data. Here's what our NCES-calibrated assumptions got right — and what they got wrong.
What We Don't Know
Small closure sample. The current real-data output contains 6 observed closures. Model comparisons are therefore exploratory, and institution-level scores should not be read as validated forecasts.
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 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.