College Closure & Financial Distress Forecaster

A survival-analysis pipeline predicting which US private nonprofit colleges will close 1 to 4 years ahead, from public data alone. It continues an interest in higher-ed economics that began with my senior thesis on test-optional admissions.

Data

The model runs on an institution-year panel of roughly 1,200 colleges from 1998 to 2024, assembled from IPEDS (via the Urban Institute API and raw NCES files), FSA/PEPS closure records, WICHE cohort projections, and Census demographics. Closure-versus-merger labels are hand-verified against roughly 20 known events, since the two look identical in most administrative data but mean very different things.

Approach

Three models: a discrete-time hazard logit for interpretable coefficients and calibrated probabilities, a Cox proportional-hazards model as a robustness check, and a class-weighted LightGBM for nonlinearities. All three are evaluated by rolling-origin backtest in which every feature respects its real publication lag. This matters more here than in most settings: IPEDS finance data arrives two years late, so a model that ignores the lag is predicting closures with data that would not have existed at decision time.

Results and caveats

The full model beats the Department of Education's own financial-responsibility score from public data alone: ROC 0.85 versus 0.74 on identical rows. Two caveats are worth stating. A two-variable benchmark, enrollment trend times tuition dependence, still wins in the extreme tail. And the demographic cliff is not yet detectable in closures through 2024 once calendar time is controlled for.

The pipeline is refreshable annually as new IPEDS data publishes, with a CLI that scores the live universe of institutions. Two of the model's 2024 top-25 highest-risk institutions, Fontbonne and Northland, closed in 2025, after the model's information window.