NCUA Credit Union Data: Validation & Reconciliation

What we checked: at the most recent quarter (Q4 2025), do our published headline figures match what NCUA released in their press release and quarterly state of the credit union system report?

This is the most important external sanity check. It confirms that our parsing of NCUA's aggregate report ties to NCUA's own headline communication.

What we checked: at every level of the loan tree, for every metric (balance, 30-59 DPD, 60+ DPD), for every quarter, does the parent value equal the sum of its children?

Tree levels checked at each quarter:

• Total = Consumer + Real Estate + Commercial
• Consumer = Credit Cards + Auto + Other Unsecured + Other Secured + Leases + Student + PAL
• Real Estate = 1-4 Family 1st + 1-4 Family Junior + Other Non-Commercial
• Commercial = Real Estate Secured + Not Real Estate Secured
• Consumer Auto = New Vehicle + Used Vehicle

Tolerance: 0.1 percent at sub-segment level, 0.5 percent at top level (allowing for small unmapped categories). The per-quarter top-level closure table (Total Loans vs Consumer + RE + Commercial) is available in the interactive tool.

NCUA classifies every loan into exactly one of three primary buckets. For our published Consumer figures to be reliable in the pre-modern era (where NCUA did not break out sub-segments), we rely on this identity to derive Consumer algebraically as Total - RE - Commercial. We prove the identity holds wherever both sides are observable.

Result: across all 16 quarters from Q1 2022 to Q4 2025 (the modern era where every sub-segment is reported), the residual is zero to four decimal places. This holds for balances, 60+ delinquency dollars, and net charge-offs.

Each headline field starts producing values at the expected NCUA reporting date. For example, Consumer Auto delinquency should first appear in Q2 2013 (when NCUA introduced the breakdown), Commercial NCO rate at Q1 2011 (first quarter where average balance is computable), etc. This confirms the leading-zero masking logic fires in the right places without over- or under-masking.

NCUA reports charge-offs and recoveries year-to-date. Within a calendar year, YTD values are cumulative and must not decrease. We scan all 50 YTD fields across all quarters and flag any drops greater than 0.5 percent or $50,000 (whichever is larger).

Result: 27 raw violations. 5 are the leases and MBL paired anomalies where NCUA's own filings have a YTD value impossibly decreasing within the same calendar year (see anomalies section). Under the publish-as-filed policy these are surfaced on the chart with explanation rather than masked. The remaining 22 are all inter-sub Commercial reclassifications - when one Commercial sub drops, another rises by close to the same amount in the same quarter. The aggregate Commercial Total YTD figures are monotonic across all 16 modern quarters.

Physical sanity. The 30-59 DPD stock plus the 60+ DPD stock combined cannot exceed the segment's outstanding balance. Each individual delinquency stock cannot exceed the balance either. Result: zero violations across roughly 1,548 segment-quarters (18 segments x 86 quarters). Delinquency rates are all within the plausible range (0 percent to 30 percent).

We independently re-compute the published net charge-off rate using only the raw YTD charge-off, recovery, and balance fields. Formula: quarterly NCO = (current YTD - prior YTD); annualized rate = quarterly NCO times 4 divided by average balance times 100. We also re-derive the legacy 30-59 stock fields (computed as 30+ minus 60+ from legacy aggregates); all 86 quarters produce a non-negative result, confirming the derivation is sound.

For every published indicator, compute the quarter-over-quarter change, then a 12-quarter rolling z-score. Flag any change with |z| greater than 3. This is the ongoing automated monitor for new anomalies in future data.

Result: 97 outliers at z >= 3, all explainable when categorized:

• 37 cluster at known event dates: COVID 2020-Q2 (forbearance), post-COVID 2022-Q2 loan growth surge, MBL/Consumer reclassification 2017-Q3, GFC peak 2008-Q4.
• The remaining 60 are tiny-book volatility (PAL, Student, Leases under $5B), Q4 batch effects in commercial reporting, Q1 reclassification resets, pre-GFC NCO ramp 2007-2008, or the Q4 2024 commercial NCO spike with Q1 2025 reversal.

All four quarters of 2025 across 12 fields cross-validated. Initial run 2026-04-22: 45 of 48 PASS - three failures, all the same single issue (the credit union count showed 4,287 for every quarter rather than 4,411 / 4,370 / 4,331 / 4,287). Reported to maker, traced to the parser broadcasting a single value to all five quarter dates, fixed at source. Re-run 2026-04-25: 45 of 45 PASS, 3 quarters NO_JSON (correctly null). Effective pass rate 100 percent.

Every balance field (Total, Auto Total/New/Used, Credit Cards, 1-4 Family Total/1st/Junior, Commercial Total/RE-Secured/Non-RE-Secured) at 9 quarters covering the last 5 years. 104 of 104 PASS. This window covers the COVID stress period and the post-COVID auto cycle.

Every field the parser supports across all 11 modern-layout PDFs (year-ends 2005-2024 plus the four quarters of 2025). 764 of 803 PASS. The 39 fails:

• 28 NCO methodology fails - the pipeline publishes the NCO rate using the FDIC-style quarterly-annualized methodology; NCUA's QCUDS uses YTD-annualized. Both valid; the chart label discloses which is in use. Intentional design so credit-union and bank rates share methodology on the same chart.
• 11 sub-1 percent restatement drifts on first-introduced segment balances (Commercial Total, Commercial RE-Secured, 1-4 Family Junior Lien) at 2017-2019. The pipeline carries the latest restated value, which is correct. Drift sub-0.5 percent in every case.

The highest-risk window: pre-2015 the maker uses label-sheet legacy aggregate logic to extend coverage back to Q4 2002. 92 of 99 PASS, all 7 fails the same NCO methodology choice. The cleanest available external evidence that the pre-GFC extension produces numbers consistent with NCUA's own quarterly publications.

*Maker correctly publishes null for intra-year quarters where the source FPR archive does not currently include the corresponding mid-year file - better null than a wrong forward-filled value. For all other fields all four quarters of 2025 cross-check PASS.

Observed gap: +3.15 percent in 2002 narrowing to +0.60 percent at Q4 2025. The trajectory matches the historical consolidation of approximately 120 non-FICU state-chartered credit unions over the 23-year window. The level gap is coverage scope, not pipeline error.

One outlier surfaced: Total Loans in Q2 2011 shows a $23 billion step-up on the Fed Z.1 side, with an offsetting reversion in Q1 2013. This is documented Fed methodology (the Z.1 reclassified "Total Loans Including Repo") - not a pipeline issue. Total Assets, the cleaner cross-check, stays at a flat ~1.3 percent gap through that window.

Methodology note: this reconciliation uses YTD-annualized NCO rate for direct comparison to NCUA's series. The published TREMOR chart uses an FDIC-style quarterly-annualized rate so credit-union and bank figures share methodology on the same chart. Both are valid and produce close but not identical numbers.

A YTD value cannot decrease from one quarter to the next within the same calendar year. We observed three pairs in the leases line where it did anyway: Q1 2007 YTD $24.0M to Q2 2007 YTD $11.0M (impossible $13M drop, deriving a -3.91% rate at Q2 2007); Q3 2013 $3.17M to Q4 2013 $2.42M; Q2 2016 $5.35M to Q3 2016 $3.03M. All three are leases-specific on a small ($0.5B-$1.5B) book. Published as filed - a chart point showing -3.91 percent with a clearly-marked explanation is more honest than a missing data point.

Through Q2 2017 NCUA reported a single aggregate MBL charge-off line (account A400T). Q3 2016 batch spike (~$190M vs ~$39-49M neighbours, +1.21% rate); Q4 2017 year-end batch (~$224M, +1.43% rate, consistent across file vintages); Q2+Q3 2018 paired impossible-YTD-decrease ($245M to $176M). Published as filed.

Q1-Q3 each year had $1-5M of incremental charge-offs, then Q4 alone showed $12-21M. On the quarterly-annualized rate chart: Q4 2016 +2.25%, Q4 2017 +1.12%, Q4 2018 +1.44% (vs ~0.4-0.6% in other quarters). Delinquency rates show no parallel rise - the Q4 spikes are charge-offs leaving the balance sheet, the loans were already overdue. Published as filed.

The Q4 2006 value was originally filed at $363M (Dec 2006 release) and revised to $580M by the Sep 2007 release. On the chart: Q3 2006 = 1.47%, Q4 2006 jumps to 2.18%, Q1 2007 settles at 1.18%. The jump has no corresponding move in any other NCUA series. We use NCUA's latest published value ($580M).

The A020B field reported $0 in Q4 2002/2003 (bucket not yet populated), then Q4 2004 $7,036M (~2.99% of Consumer), Q4 2005 $5,883M (~2.15%), Q4 2006 $5,595M (1.75%, first consistent reading). Both 2004 and 2005 implausibly exceed the eventual GFC peak with flat 60+ DPD; FDIC banking data shows no parallel spike. Most likely a first-wave reporting artifact as NCUA stabilized the new field. Published as filed.

Account A386B first values: Q3 2017 $15.8B, Q4 2017 $11.7B, Q1 2018 $4.8B, then $3-4B. The other two new RE sub-codes grew smoothly through the same transition, consistent with credit unions initially booking loans into the new "other" bucket and NCUA refining the definition. Published as filed; parent Real Estate aggregate goes back to natural sum-of-three-subs (see Layer 1 Callout 2).

The Dec 2024 Aggregate FPR workbook has the "Count of CU :" label cell but an empty value cell - NCUA published the formula without saving the cached value. Q4 2024 Count of CU is shown as null rather than a wrong number; follow-up is to acquire a re-saved file or apply a press-release override (~4,493). Other Dec 2024 fields are unaffected.

Real Estate Q1 2022 - Q2 2023: post-COVID recoveries exceeded new charge-offs as the foreclosure moratorium ran off (rates within +/- 0.01%, visually flat). Consumer Leases Q2/Q4 2012, Q4 2020, Q4 2021: small swings on a tiny book (rates -0.04 to -0.27%). These are real NCUA data and would be incorrect to mask.

Before Q1 2011, Member Business Loans were embedded inside legacy Real Estate aggregate codes. NCUA introduced A400T (Total MBL) in Q1 2011. On the charts: Real Estate $338.4B (Q4 2010) to $300.2B (Q1 2011); Commercial $0 to $37.5B; Total Loans $564.7B to $559.9B (normal small movement). The two steps almost exactly offset.

NCUA retired several legacy aggregate codes and introduced granular sub-segment codes. On the charts at Q3 2017: Consumer $415.8B to $466.1B (+$50B, of which $37.7B is the brand-new A698C "Other Secured Non-RE" code); Real Estate $424.7B to $409.1B (-$15.6B); Commercial $72.5B to $61.8B (-$10.7B). Total Loans $913.0B to $937.0B is normal +$24B organic growth - the bucket shuffles net to zero at the Total level.

From Q1 2022 NCUA reports nine commercial sub-categories. Layer 1 found 22 cases where a single sub-category's YTD value decreased - but every time one drops, another rises by close to the same amount. The aggregate Commercial Total YTD is monotonic across all 16 modern quarters. NCUA reclassifies loans between commercial sub-categories as new information becomes available.

NCUA's Total 30-59 DPD field reported a literal $0 in Q4 2002 and Q4 2003 - not because delinquency was zero, but because the bucket wasn't yet a reported field. We mask the leading run of zeros for several 30-59 codes so the chart doesn't show a misleading 0.00 percent rate at the start of the series. This is a schema-placeholder zero, treated differently from the publish-as-filed policy applied to NCUA's real values.

NCUA introduced separate delinquency reporting for new vehicle and used vehicle loans beginning Q2 2013. Before that, auto loans were reported only as part of broader Consumer delinquency aggregates. We carry Consumer Auto delinquency from Q2 2013 forward and leave earlier quarters as null rather than fabricating an estimate.