Line-item tags mapped to balance-sheet total concepts — `standard_concept == 'Assets'` returns 3 rows on a bank filing

[DerMayer1]

Line-item tags are mapped to balance-sheet total concepts, so a standardized statement returns several rows claiming to be the total. Affects standard_concept on any statement dataframe.

Version 5.43.0.

Repro

from edgar import Company

df = Company('CFG').get_financials().balance_sheet().to_dataframe()
df[df['standard_concept'] == 'Assets'][['concept', 'label', '2025-12-31']]

Citizens Financial Group, FY2025 — three rows:

concept	label	2025-12-31
us-gaap_DebtSecuritiesHeldToMaturityExcludingAccruedInterestAfterAllowanceForCreditLoss	Debt securities held to maturity	7,933,000,000
us-gaap_BankOwnedLifeInsurance	Bank-owned life insurance	3,441,000,000
us-gaap_Assets	TOTAL ASSETS	226,351,000,000

Selecting on the standardized concept and taking the first row reports total assets as $7.9B instead of $226.4B.

No network needed:

from edgar.xbrl.standardization.core import MappingStore
MappingStore().get_standard_concept('us-gaap_BankOwnedLifeInsurance')   # 'Total Assets'
MappingStore().get_standard_concept('ifrs-full_IntangibleAssetsAndGoodwill')  # 'Total Non-Current Assets'

edgar/xbrl/standardization/reverse_index.py. 53 index entries sit in section: 'Totals' with is_total: True. Most are line items:

Total Assets            Assets, AssetsNet, ifrs:Assets
                        + AccruedInvestmentIncomeReceivable, BankOwnedLifeInsurance,
                          CashCashEquivalentsAndFederalFundsSold, FederalFundsSold,
                          FederalHomeLoanBankStock, InvestmentOwnedAtFairValue,
                          DebtSecuritiesHeldToMaturityExcludingAccruedInterest...

Total Liabilities       Liabilities, ifrs:Liabilities
                        + AdvancesFromFederalHomeLoanBanks, DueToRelatedPartiesCurrentAndNoncurrent,
                          ManagementFeePayable, OtherBorrowings, SecuredDebt,
                          SecuritiesSoldUnderAgreementsToRepurchase, SubordinatedDebt,
                          UnearnedPremiums, WarrantLiability,
                          SharesSubjectToMandatoryRedemptionSettlementTermsFairValueOfShares

Total Current Liabilities   LiabilitiesCurrent, CurrentLiabilities
                        + CurrentBorrowingsAndCurrentPortionOfNoncurrentBorrowings,
                          CurrentDerivativeFinancialLiabilities, CurrentPortionOfLongtermBorrowings,
                          CurrentProvisions, CurrentTaxLiabilitiesCurrent,
                          LiabilitiesIncludedInDisposalGroupsClassifiedAsHeldForSale,
                          OtherCurrentFinancialLiabilities, TemporaryEquityLiquidationPreference,
                          TradeAndOtherCurrentPayablesToRelatedParties

Total Current Assets    AssetsCurrent, CurrentAssets
                        + CurrentDerivativeFinancialAssets, CurrentPrepaymentsAndOtherCurrentAssets,
                          CurrentTaxAssetsCurrent, CurrentTradeReceivables, OtherCurrentReceivables,
                          CurrentAssetsOtherThanAssetsOrDisposalGroups...

Total Non-Current Assets    AssetsNoncurrent, NoncurrentAssets, ifrs:NoncurrentAssets
                        + IntangibleAssetsAndGoodwill, NoncurrentDerivativeFinancialAssets

The concentration is in bank and insurer line items (federal funds sold, FHLB stock, BOLI, unearned premiums, subordinated debt), which is why it doesn't show on a typical industrial filer. us-gaap_Goodwill maps correctly to Goodwill; the mis-mapped set is specific.

Separately, _normalize_tag (reverse_index.py:188) strips ifrs-full_ / ifrs-full: before lookup, so IFRS tags resolve against the GAAP index by bare name. On VEON's 20-F, ifrs-full_IntangibleAssetsAndGoodwill and ifrs-full_CurrentProvisions are labeled Total Non-Current Assets and Total Current Liabilities. VEON has 24 of 111 rows standardized, Klarna 16 of 81, so IFRS coverage is thin to begin with, the wrong mappings are the part that's actively misleading rather than just missing.

Random sample of 60 tickers from get_cik_tickers(), checking Assets == Liabilities + Equity per rendered period. The identity holds everywhere (0 mismatches over 98 periods, plus 49 large caps) because it reads raw concepts, the defect is only in the standardization layer.

Is gaap_taxonomy_mapping.csv the right place to fix this, or the index generation? The section: 'Totals' bucket looks like it's doing double duty, "this tag belongs to the totals region of the balance sheet" vs "this tag is the total", and is_total: True on a component is what makes it selectable as one. If the intended fix is in the source CSV I can't regenerate the index from the repo, since data/xbrl-mappings/ isn't checked in.

Happy to send a PR once you've said which layer it belongs in. A cheap guard in the meantime would be refusing is_total: True for a tag whose standard_tags is a *Total concept it isn't the canonical tag for, but that's a patch on the output.

[dgunning]

Confirmed and root-caused. Thanks for the no-network repro and for pointing at reverse_index.py — that shortened the investigation considerably.

One correction to the diagnosis: the culprit is not is_total. us-gaap_BankOwnedLifeInsurance already carries is_total: false and still resolves to Total Assets, so filtering on that flag would not fix it. The real problem is confidence. gaap_mappings.json assigns standard_tags: ["Assets"] with display_name: "Total Assets" to low-confidence fuzzy matches, and the consumer surfaces them as authoritative with no distinction from the canonical mapping.

The distribution is cleanly bimodal, which is what makes this tractable:

• 124 tags map to a Total* display name
• 20 at confidence ≥ 0.9 — the canonical ones (Assets, Liabilities, AssetsCurrent all sit at exactly 1.000)
• 59 at confidence < 0.5 — the fuzzy matches, clustered around 0.307

So a confidence guard in the consumer removes all 59 bad mappings in one change without touching the good ones.

Your IFRS example is a second, independent bug: _normalize_tag (reverse_index.py:188) strips the ifrs-full_ prefix before lookup, so IFRS tags resolve against the GAAP index by bare name. That is how ifrs-full_IntangibleAssetsAndGoodwill lands on Total Non-Current Assets.

One wrinkle on where the fix goes: gaap_mappings.json is generated upstream (_metadata.source: "edgar-storage StandardizedConcepts", v3.0.0, built from 32,240 filings). Hand-edits to it survive in-repo but are not round-tripped, so they would be lost on the next regeneration. The plan is therefore a confidence guard in the consumer — which is where the authority judgement belongs anyway — plus an upstream issue against the generator and a regression test pinned to your CFG case.

Tracked as edgartools-gnx5 (P1).

[DerMayer1]

You're right about is_total, and my suggested guard would have missed the two worst cases. BankOwnedLifeInsurance is is_total: false at confidence 0.319, FederalFundsSold the same at 0.309, so filtering on the flag leaves both resolving to Total Assets. Confidence is the right discriminator.

I went and checked where the threshold actually needs to sit. Sampled 65 filings (6 banks and insurers picked deliberately, 59 random from get_cik_tickers()) and looked for rows whose resolved standard_concept collides with the canonical total row on the same statement. 29 colliding rows, 12 distinct tags, and every single one sits below 0.5, in a tight band from 0.307 to 0.319. Nothing between 0.5 and 0.9 collided at all.

One thing that surprised me, and it works in favour of the cut: the confidence distribution isn't quite bimodal. 124 tags carry a Total* display name, 20 at ≥0.9, 3 in 0.7-0.9, but 42 sit in 0.5-0.7. I expected some of those to be the same bug and they aren't, because context disambiguation catches them before they surface. CommonStocksIncludingAdditionalPaidInCapital is the clearest example: display name "Total Stockholders' Equity" at exactly 0.5, but on AAPL and MSFT it resolves to CommonEquity, which is correct. So the middle band looks alarming in the index and is harmless in the output. A cut at 0.5 clears everything I could actually reproduce.

If it's useful for the regression test, here's what showed up across 15 banks and insurers, all in that 0.307-0.319 band:

tag	resolves to	filings
DebtSecuritiesHeldToMaturityExcludingAccruedInterestAfterAllowanceForCreditLoss	Assets	9 (BAC, C, CFG, JPM, PNC, ...)
AccruedInvestmentIncomeReceivable	Assets	5 (AIG, ALL, MET, PRU, TRV)
UnearnedPremiums	Liabilities	3
SubordinatedDebt	Liabilities	2
AdvancesFromFederalHomeLoanBanks	Liabilities	2
BankOwnedLifeInsurance, SecuredDebt, SecuritiesSoldUnderAgreementsToRepurchase, OtherBorrowings	Assets / Liabilities	1 each

CFG is a good pin because it carries two at once, but held-to-maturity securities is the one with real reach across the big banks.

Happy to send the consumer guard plus a fixture over those tags if that helps. Say the word if you already have it in flight and I'll stay out of the way.