Final Lock Note
Status: locked partial domain-specific extension; hardened FASB × Sector 2 private nonprofit four-year closure prediction; not production.
Summary
The Higher-Ed A/G Research Extension is a research-grade public-data study. It applies a two-layer geometry — Ground (institutional finance) and Appearance (institutional performance and continuation signals) — to U.S. higher-education institutions, using IPEDS data 2014–2023. The validated lane is intentionally narrow: FASB-filing private nonprofit four-year institutions, evaluated for closure within five years.
Validated lane
- Accounting family: FASB (private nonprofit financial reporting).
- HD sector: Sector 2 (private nonprofit four-year).
- Target: closure within five years (forward window from observation year T; closure year ∈ {T+1, …, T+5}).
- Primary geometry: joint-weak G+A (institutions whose Ground and Appearance scores are both below the within-stratum median).
Key findings
- On the locked test cohort (test years 2017–2018), joint-weak G+A predicts closure within five years at AUROC ≈ 0.864.
- Institution-cluster bootstrap 95 % CI ≈ [0.806, 0.914] — almost identical to the original analytic interval at the lower bound.
- The signal survives four sliding temporal folds at AUROC 0.77–0.86. Within-A/G ordering (joint-weak above G-only above A-only above divergence) holds in every fold.
- The geometry has a higher AUROC point estimate than every tested single-variable baseline. Cluster-bootstrap paired-difference 95 % CI vs the strongest baseline is [+0.002, +0.138]; about 98 % of bootstrap samples are positive. This is a higher point estimate, not a definitive statistical-superiority claim.
- For the secondary enrollment-collapse target, simple low six-year graduation rate alone has a higher AUROC point estimate than the A/G geometries. The framework does not beat baselines for that target.
- Closure signal is detectable across the full T-1 through T-5 horizon — a multi-year early-warning pattern, not a coincident T-1 alarm.
What is not validated
- Public-sector four-year institutions (closure events too rare in the panel window).
- For-profit institutions (endowment-cushion is structurally absent; admissions/graduation reporting is sparser).
- Two-year and community-college institutions (the four-year-cohort Appearance layer does not apply).
- Enrollment-collapse targets (baseline-dominated).
- Head-to-head comparison against the federal Financial Responsibility composite on the locked test cohort (data-coverage gap).
- Heightened Cash Monitoring as an anchor (source-blocked in this run).
Limitations
Read with care. AUROC point estimates and CIs are summary numbers from a single time-block split (Phase 15) plus four sliding folds (Phase 19). Joint-weak’s edge over the strongest single-variable baseline is small at the cluster-bootstrap paired-difference lower bound (+0.002). Top-decile flagging captures only a subset (about 13–27 %) of pre-closure rows across the T-1 through T-5 horizon. This is a research artifact in a narrow lane, not a closure-prediction product.
Reading order
- Claim Boundary Note — what may and may not be claimed.
- Phase 19 Hardening Decision — the integrated decision document.
- Cluster Bootstrap, Temporal Folds, Lead-Time — the three positive hardening tracks.
- ED Financial Responsibility Baseline and HCM Anchor Note — the partially-runnable and source-blocked tracks.