U.S. College Closure Risk: The Full Analysis

The Findings

Key Results

The Story Behind the Numbers

The study asks whether public institutional records can support a useful early-warning screen, and how much model complexity that decision warrants. It began with a synthetic cohort to build the workflow and stress-test the fidelity ladder.

We asked: can you predict which institutions will close next? And more importantly: how much model complexity do you actually need?

The later public-data run changed the answer. It retained 1,516 institutions and only 6 observed closures, compared with 244 synthetic closures. Model performance from the two cohorts cannot be treated as interchangeable.

The defensible conclusion is that no model here is ready to guide a real closure decision. The work still has value as an ADM case study: validation correctly forced a lower-confidence conclusion and identified data coverage, not added complexity, as the next bottleneck.

What We Used

All data sources are publicly available. Pipeline ingests real College Scorecard and IPEDS data.

Three Tiers of Analysis

Tier 1: Screening Rule (3 variables)

The simplest possible model. An institution is flagged as “at risk” if enrollment has declined for 3+ consecutive years AND it derives >85% of revenue from tuition AND has fewer than 90 days of cash reserves.

The rule is easy to compute and audit, but the current public output reports 0% recall, 0% precision, and 59 false positives. Interpretability makes it a useful prototype, not a validated warning system.

Tier 2: Multi-Variable Model (12 variables)

Adds regional demographic headwinds (Census-calibrated birth rate projections), completion rates, retention rates, default rates, market competition (schools per 18-year-old in the state), and online education share. Logistic regression reaches 40% recall in the current cross-validation output, with an AUC of 0.849.

Geography and demographics remain plausible features, but the six observed closures are spread across four regions. The current output is too sparse to show that these additions improve a practical closure decision.

Tier 3: Gradient Boosting Classifier (17 variables)

A gradient boosting classifier trained on the retained feature set. The current output reports 0% recall at the selected threshold and an AUC of 0.901.

The AUC suggests a possible ranking signal, but the threshold result and tiny positive class prevent a credible claim of calibrated risk or useful recall. Feature rankings should be treated as hypotheses for a larger validation set.

Six Questions

Q1: Which Colleges Are Most at Risk?

Financial stress score using enrollment trend, tuition dependency, and cash reserves. The 3-variable rule flags institutions with declining enrollment (>2%/yr), high tuition dependency (>85%), and low reserves (<90 days cash). In the current output, the selected rule flags none of the six observed closures.

Q2: How Much Does Geography Matter?

The West has 2 observed closures in 180 institutions; the Southeast also has 2. Regional rankings are descriptive and too sparse for a demographic-causation claim.

Q3: When Does Decline Become Fatal?

Enrollment below 900 students combined with 3 or more years of decline identifies 93 institutions; 1 was observed closed, a 1.1% descriptive rate.

Q4: Did Online Colleges Win or Lose?

No closures were observed among the 74 primarily-online or 10 fully-online institutions in this cohort. That is too little evidence to estimate an online-by-sector interaction.

Q5: Can ML Beat the Rule?

Logistic regression reports 40% recall; the stress rule and gradient boosting each report 0%. With six observed closures, the comparison is a warning about uncertainty rather than a model-selection result.

Q6: How Do We Know the Models Work?

The synthetic cohort had a 5.6% closure rate; the real-data cohort had 0.4%. The mismatch shows why synthetic development results cannot serve as real-world validation.

What We Don't Know

Six observed closures is not a deployment sample. Cross-validation folds contain roughly one positive case each, so recall and precision can swing sharply with a single classification.

Mergers aren't closures. Several institutions counted as “closed” in NCES data were actually merged or acquired. Their students were absorbed, not displaced. The closure impact is overstated for these cases.

We can't model governance. Some closures result from leadership failures, financial fraud, or accreditation loss — events that don't appear in financial data until it's too late. No model can predict a board's decision to close.

How We Built This

This study follows the Analysis Driven Modeling (ADM) framework: start with the question, match fidelity to the decision, encode what you know, learn the rest.

The question: Which U.S. colleges are at risk of closing, and how much model complexity do you need to answer that question?

The approach: We started with the simplest possible model (3 financial ratios) and escalated fidelity only when the simpler model left meaningful signal on the table. Each tier adds variables and complexity — and we measured whether the added complexity was worth the cost.

Validation: The public output compares model behavior against six observed closure labels in the retained real-data cohort. That is an exploratory check, not external validation, and it does not establish calibrated institution-level risk.

Techniques used: Rule-based threshold model, logistic regression with domain features, gradient boosting classifier (sklearn), Kaplan-Meier survival analysis. No neural networks — the dataset is too small to justify them.

Python pipeline: ~2,000 lines. All source code and data available on request.