Skip to content

What 2.1 Billion Filings Reveal About the Bank Regulatory Rulebook: Empirical Findings from a Verified Data Dictionary

A researcher-facing summary of the empirical findings from the v8–v10 validation campaigns. Every number below traces to Empirical Validation, the Relationship Registry, and the official edit sources cited there.

1. Setting — what was tested

A data dictionary usually asserts that "assets equal liabilities plus equity" and stops. This project tests every such assertion as an aggregate query against the full history of real bank filings, then publishes the pass rate, the observation count, and an honest verdict for each.

What was tested:

  • The official rules. The FFIEC Call Report edit checks and calculation-linkbase identities, taken directly from the regulators' own machine-readable formula linkbases (the CDR XBRL taxonomy, cdr.ffiec.gov) — 3,615 edits per form plus 243 calculation-linkbase identities across forms FFIEC 031 / 041 / 051 — and the FR Y-9C edit checklist, parsed the same way. To these were added the dictionary's curated reconciliation identities and component hierarchies, and (in v10) the UBPR derivation formulas.
  • The corpus. These rules were evaluated against the full bulk filing record: 208 million reported cell-values of FR Y-9C (bulk BHCF) data, every quarter 1986 Q3 – 2025 Q4, 13,668 holding companies; and 1.917 billion rows of Call Report filings, 2001 Q1 – 2026 Q1.
  • The registry. The result is a single 7,539-relationship registry, every row carrying a status and — where machine-testable — its empirical observation count and pass rate. Zero rows are left pending or unexplained. Of these, 2,464 Call relationships and the FR Y-9C set were machine-evaluated; a 16-concept cross-source concordance and 31 UBPR ratios were validated independently.

A relationship "passes" for a filing when both sides agree within max($1,000, 0.1%) of the larger side — enough to absorb thousands-rounding without masking a real break. Every non-passing row was adjudicated against the official edit text, the field specification, and the MDRM — never silently dropped.

2. Finding I — official rules that don't hold in the filed data

Eight official edits do not hold even on the correct window, population, and all-components-reported subset. They are kept deliberately as research findings about how the official edit set behaves against real public data, classified OFFICIAL_EDIT_UNMET. The two headline cases:

(I-a) A phantom decomposition. The official calculation linkbase decomposes the representation-and-warranty reserve RCONM288 into RCONL191 + RCONL192. But L191 and L192 have zero observations warehouse-wide in the public bulk data, while M288 is reported directly in 51,992 filings. The decomposition is not empirically derivable from anything public; the edit fails at 0.828 and is documented, not enforced (registry REL_7424). A second, RCON-side variant of the same family (REL_5161, the same M288 = L191 + L192 reserve, edit R6230.5551) was surfaced by the v10 conditional engine and is the 8th finding — pass 0.937, same root cause. Reproduce: over all Call filings, compare RCONM288 against RCONL191 + RCONL192; note that L191/L192 are never populated, so the sum collapses to zero while the parent is reported.

(I-b) A securitization bound that mostly fails among the population it's about. An RC-R Part II securitization bound (REL_5882, edit R7020.6507) holds at a headline 0.983 — but that rate is dominated by zero-filers (banks with no securitization, for whom both sides are 0). Among active securitizers — the population the edit is actually about, n = 3,493 — it holds at only 0.452. That is a real, substantial explained-violation population, not a rounding tail. Reproduce: restrict to bank-quarters where the securitization components are non-zero, then test the bound; the pass rate falls from 0.983 to 0.452.

The remaining six are soft "should-be" quality edits whose violation rates are themselves the finding: an RC-O brokered-deposit bound at 0.714 (REL_5399, R6770.5451) and a family of RI-C / RC-C CECL loan-category bounds at 0.657 – 0.937 (REL_3757/3759/3760/3763, edits R2006.7256/7258/7259/7262). All eight rows, with full expressions and per-row empirical_n / empirical_pass_rate, are in the Relationship Registry filtered to OFFICIAL_EDIT_UNMET.

3. Finding II — corrected cross-source definitional mappings

For 16 core concepts the Call Report value was compared against the FDIC's independently-collected Statistics on Depository Institutions (SDI), bank-quarter by bank-quarter, over up to 667,551 bank-quarters. All 16 concord at 99.5 – 99.97%. More usefully, the concordance surfaced two mappings that the naïve crosswalk gets wrong:

(II-a) Total equity → Call item 3210, not G105. FDIC EQ maps to RCFD/RCON3210 (total bank equity, excluding minority interest). The minority-interest-inclusive total G105 scored only 98.6% against EQ; 3210 scores 99.72% (n = 666,679). The correct FDIC↔Call equity mapping therefore excludes minority interest.

(II-b) Net loans → 2122 − 3123, not B529. FDIC LNLSNET maps to loans net of unearned income (2122) minus the allowance (3123), not the Schedule RC-C item-12 net total B529. The B529 basis scored only 77.8%; the 2122 − 3123 expression scores 99.89% (n = 656,563). The two "net loans" totals are defined differently, and the SDI uses the 2122 − 3123 basis.

Why it matters: anyone joining FDIC SDI to Call Report data on "equity" or "net loans" with the obvious code will inherit a 1.4- or 22-point definitional mismatch that looks like noise but is structural. Both corrections are recorded in json/cross_form_mapping.json under total_equity and total_loans; the full 16-concept table is in _rebuild/empirical/RESULTS_cross_source.csv.

4. Finding III — the rulebook's 25-year growth

The official Call Report edit set was parsed from 30 reporting cycles (every year-end 2001–2024 plus all 2025–2026 quarters, forms 031/041/051), one machine-readable lifetime per edit label. Headline numbers (Edit History, csv/EDIT_HISTORY.csv):

  • 15,622 distinct edit labels observed, 2001–2026.
  • 4,759 retired before the current (2026 Q1) cycle.
  • 1,212 introduced in 2020 or later.
  • Edit count grew from ~2,450 (2001 year-end) to ~11,000 (2025 Q1) — the rulebook roughly quintupled in 25 years, with quality ("should"-hold) edits outnumbering validity ("must"-hold) edits about 2.5 : 1 throughout.

The eras are visible in the data: a steady state (2001–2007), a crisis buildout in past-due / securitization / fair-value reporting (2008–2013), the Basel III spike (2014–2019 — the single largest event, validity edits jumping from ~2,100 to ~3,300 while hundreds of pre-Basel capital edits retire simultaneously), the CECL and pandemic turnover (2020–2022 — ASU 2016-13 allowance edits replace incurred-loss edits; short-lived PPP items appear and depart), and a refinement plateau (2023–2026, where change shifts to expression revisions within continuing labels).

5. Finding IV — UBPR construction conventions, pinned empirically

The Uniform Bank Performance Report (UBPR) turns each bank's raw Call Report into derived ratios. The official CDR XBRL UBPR taxonomy v181 and User's Guide were acquired, 4,207 derivation formulas parsed with zero silent drops, and the 31 headline ratios recomputed from Call/UBPR inputs and compared to the real published UBPR value over 2015 Q1 – 2026 Q1 (~44,000 bank-quarters per ratio). All 31 validated at 99.77 – 100% (e.g. ROA UBPRE013 at 99.79%, ROE UBPRE630 at 99.77%, net interest margin UBPRE018 at 99.77%, efficiency UBPRE088 at 100.0%). Calibrating the validation pinned three construction conventions that a naïve reading gets wrong (UBPR Guide):

  1. Units. Call dollar items (RIAD/RCON/…) are in thousands; UBPR average/level concepts (e.g. UBPRD659) are in dollars. A ratio mixing the two must divide the UBPR value by 1,000.
  2. Annualization. Year-to-date flows annualize by ×(4/q) — Q1 ×4, Q2 ×2, Q3 ×4⁄3, Q4 ×1.
  3. Averages as published. Denominators like UBPRD659 (average total assets) are taken as published year-to-date averages of the Schedule RC-K quarterly balances, not re-derived.

Caption-mislabel warning. The taxonomy's presentation captions mislabel some concepts; the User's Guide is authoritative. UBPR7408 and UBPRE635 read like capital-adequacy measures but are 12-month growth rates ("Tier One Capital 12-month growth rate"). They validate as parsed but must not be read as adequacy ratios. Relatedly, the capital ratios UBPRD486/487/488 are regulator passthroughs stored 2-decimal-rounded as a fraction — validated as the documented ×100 relationship within a ±0.5 pp band, not recomputed from RC-R primitives.

6. Methods in brief

  • Tolerance. A relationship passes a filing when the two sides agree within max($1,000, 0.1%) of the larger side — absorbing thousands-rounding without masking real breaks.
  • Windows and populations. Each edit is tested on the window and sub-population where its codes are actually collected. Where a headline rate is dominated by trivial zero-filers (e.g. the securitization bound), the rate among the active population is reported alongside it.
  • Adjudication to zero unexplained. Every non-passing row was investigated against the official edit text, field spec, and MDRM and assigned a defended verdict — CONFIRMED, CONFIRMED_CURRENT, QUALITY_TOLERANCE, DATA_GAP (sparse child line-items, identity correct), NOT_IN_BULK, CONDITION_NEVER_MET, or OFFICIAL_EDIT_UNMET — never dropped. Notably, the RC-O probability-of- default-bucket family (CONDITION_NEVER_MET, 48 rows) references 280 codes that are confidential supervisory items, verified absent from all public bulk data, so its triggering condition is never met in any public filing.
  • Independent verification, including of our own tester. Adjudication repeatedly caught harness defects rather than data errors — a dropped distributed-negation (A − (B + C) read as A + B + C), a silently dropped multiplication (X = Y × 0 read as X = Y), and a registry expression truncated at ~588 characters. Each was traced to its source and repaired against the official edit text before any verdict was recorded.

Full method, verdict taxonomy, and the complete results distribution: Empirical Validation and the Relationship Registry data page.

7. How to reproduce, and how to cite

Everything needed to reproduce the verdicts is public:

  • The Relationship Registry — every relationship, its expression, source (official edit id and/or calculation-linkbase parent), status, observation count (empirical_n), pass rate (empirical_pass_rate), and test window — is in csv/RELATIONSHIP_REGISTRY.csv.
  • The bulk FR Y-9C (BHCF) and Call Report filings are published by the Federal Reserve / FFIEC and are downloadable and queryable via the companion data package, FreeNIC.
  • The official edits are the FFIEC Call Report edit checks and calculation linkbases in the CDR XBRL taxonomy (cdr.ffiec.gov) and the FR Y-9C edit checklist.
  • scripts/validate_reconciliation.py --scope {y9c,call,all,ubpr} evaluates the registry against any supplied filing. Take any CONFIRMED row, evaluate its expression against the bulk filings with the max($1,000, 0.1%) tolerance, and you will recover its published pass rate.

Citation.

Bank Regulatory Data Dictionary — Empirical Validation (v10.0). The validated relationship registry and edit-history dataset. https://github.com/andenick/bank-data-dictionary

Data companion: FreeNIC — bulk FR Y-9C and Call Report filings. https://github.com/andenick/FreeNIC