Model Card · NCUA 5300
A quarterly model that ranks every federally insured U.S. credit union by its risk of disappearing — merging or liquidating — within roughly the next year, learned from seven years of NCUA 5300 Call Report filings.
Scores are a relative risk ranking, not calibrated probabilities. Read the limitations before acting on an individual score.
For each active credit union in a given quarter Q, the model estimates the likelihood that the institution stops filing within the next four quarters — the observable signature of a merger or liquidation. The label starts from disappearance in the Call Report corpus — a CU that files in Q but is absent thereafter — and is then cross-checked against NCUA's official merger and liquidation records (see limitations). Confirmed re-charters (a CU continuing under a new number) are removed as false positives.
Output is a score in [0, 1], bucketed into three bands by rank
within the scored population:
The pipeline is a deterministic sequence, rebuilt each quarter as new filings publish:
AcctDesc.txt at ingest, never hard-coded.disappeared_after = its last-seen date.HistGradientBoosting) on a time-based split — older quarters train, recent quarters validate — with balanced class weights for the rare event.Permutation importance on the validation set (contribution to PR-AUC). The signal is economically coherent: small, shrinking, thinly-capitalized, unprofitable credit unions are the ones that get absorbed.
| Feature | Reads as | Importance |
|---|---|---|
| roa | Return on assets — profitability | |
| log_assets | Institution size | |
| net_worth_ratio | Capital cushion vs. assets | |
| total_members | Membership scale | |
| assets_chg_4q | Asset growth, trailing year | |
| members_chg_4q | Membership trend, trailing year |
Disappearance is rare — about 4.7% of CU-quarters in the validation window. Against that base rate, PR-AUC of 0.30 is a 6.5× lift, and ROC-AUC of 0.795 indicates solid rank-ordering. Accuracy is not reported: a model that called everything "safe" would score 95% and be useless.
At a working threshold that flags the riskiest 5% of institutions, the model captures ~32% of actual disappearances at ~30% precision — more than six times better than chance, and a practical shortlist for an examiner or analyst to review. Cross-checking labels against NCUA records (v2) lifted PR-AUC from 0.29 and recall from ~30% by removing false-positive re-charters from training.
Source: NCUA 5300 Call Report
quarterly files, 2019-Q1 through 2026-Q1 — 29 quarters, ~142,000
CU-quarter observations across 5,569 distinct credit unions. Core
metrics (assets, members, shares) resolved for 100% of filers; the net-worth
ratio uses NCUA's reported figure (ACCT_998) where present.
Labels are reconciled against the NCUA Insurance Report of Activity
(approved mergers, with merging and continuing charter numbers) and the
conservatorships & liquidations list. Everything is reproducible from raw
files via the pipeline.
Labels assume corpus completeness through the latest ingested quarter — a "negative" is trusted only where the full forward horizon is observed, and disappearance detection relies on every published quarter being present.