The Five Ways a Balance Sheet Fails the Identity Check
Run A = L + E over every filing you hold and the failures stop looking random inside a day.
They cluster. Five mechanisms account for nearly all of them, and they are not variations on one problem. Two are filings that balance perfectly well once read on the terms they were written on. One is noise, one is a gap in your own extraction, one is a company that filed something wrong. Four different responses.
The response that ruins the exercise is to take whichever candidate figure comes closest to closing the gap and return it. That turns a wrong number into a wrong number nobody can audit: the discrepancy that would have told you it was wrong is the thing you just closed.
1. A tag the extractor did not read
The components do not add up because one is missing from your data rather than from the filing: a company-specific extension tag, a concept from a taxonomy version your mapping predates, or an ordinary us-gaap element nobody thought to map. The filing is complete. Your copy of it is not.
Detection. The gap is far larger than rounding, and the filing's raw facts contain a line that never reached your reconstructed right-hand side. The decisive version of that test needs one more field, and it gets its own section below: the same field separates this category from the last one.
Flag it as a missing tag, name the ticker, and do not pick a substitute. A near-enough figure in place of an unread one is wrong, plausible, and erases the evidence that anything was unread.
BlackRock is the ordinary example in my data: a redeemable noncontrolling interest in a form my mapping still does not reach, so liabilities plus equity falls short and the filing is flagged rather than filled in. Note that the line I am failing to read is itself a mezzanine line. The category is about whose fault a gap is, not about which row it sits on.
2. Noncontrolling interests
A parent that owns most of a subsidiary consolidates all of it. Every asset and liability lands in the parent's totals at full value, because the parent controls them. The slice it does not own is not netted out of assets; it sits on the claims side as a noncontrolling interest.
So a consolidated filing is written on A = L + E + NCI, and
testing A = L + E against parent-only equity misses by exactly the
minority stake.
Detection. The gap matches the reported
MinorityInterest or
StockholdersEquityAttributableToNoncontrollingInterest to within
the tolerance. Not approximately. Exactly, which is what makes this one safe to
act on.
Reconcile on A = L + E + NCI and record that basis. This is a pass. The filing was never wrong.
Most filers hand you the answer already summed:
StockholdersEquityIncludingPortionAttributableToNoncontrollingInterest
carries the minority interest inside it, and a filing that publishes it
balances plainly. So the category only fires where a filer reports parent
equity and the NCI as two lines with no combined total, which in the universe I
cover is rare enough to count on one hand. It is nearly empty because the
common case is settled a step earlier.
3. Mezzanine equity
Redeemable preferred stock, redeemable noncontrolling interests and shares subject to possible redemption sit between the liabilities and equity sections, in a band of their own. The instrument can be required to be redeemed, which is debt-like, while carrying none of the fixed obligation that makes debt debt. Genuinely neither, so it is shown as neither.
The identity that filing was written on is
A = L + E + Mezzanine.
Detection. The gap matches the filer's temporary equity or
redeemable preferred line:
TemporaryEquityCarryingAmountIncludingPortionAttributableToNoncontrollingInterests
and its parent-only and plain variants,
RedeemablePreferredStockCarryingAmount, or
RedeemableNoncontrollingInterestEquityCarryingAmount.
Reconcile with the mezzanine included, and say so. Also a pass.
One trap sits inside that detection. Temporary equity is the section total and redeemable preferred is a component of it, so adding both overshoots and breaks a filing that balanced. Prefer the total. Never add a section to its own parts.
KKR is the live example: a filing that does not close on A = L + E and does close once the redeemable block is added, with the drawing naming the basis. It matters most, though, on the blank-cheque company. A pre-merger SPAC holds a trust account as its only real asset and carries most of the shareholder money as Class A stock subject to possible redemption, which is temporary equity. Drop the mezzanine line there and the arithmetic does not miss by a little: liabilities plus permanent equity comes out a rounding error beside total assets, reading like an extractor that has failed outright rather than one absent term.
The classifier, in full
Three of the five are in view, so here is the whole decision: the three totals, the two extra terms where the filer published them, and one field the last two categories turn on.
TOLERANCE = 0.005 # half a percent of assets
ROUNDING_BAND = 0.01 # and one percent is the outer edge of noise
def classify(assets, liabilities, equity,
nci=None, mezzanine=None, stated_rhs=None):
# Returns (verdict, basis). A term is admitted only where the filer
# published it, and only after the plain sum has already failed:
# adding a term to a filing that balances breaks one that was right.
if assets <= 0:
return "not testable", None
def gap(extra=0.0):
return abs(assets - (liabilities + equity + extra)) / assets
if gap() <= TOLERANCE:
return "balances", None
# The two terms the plain identity does not carry. Try each and both,
# and let the TIGHTEST fit win rather than the first to clear: where
# two bases close, the tighter one is the shape the filer used.
bases = []
if nci:
bases.append(("noncontrolling interests", nci))
if mezzanine:
bases.append(("mezzanine equity", mezzanine))
if nci and mezzanine:
bases.append(("both", nci + mezzanine))
closed = sorted((gap(x), name) for name, x in bases if gap(x) <= TOLERANCE)
if closed:
return "balances", closed[0][1]
if gap() <= ROUNDING_BAND:
return "rounding", None
# Still open, and whose fault it is turns on one field: the total the
# filer printed at the foot of the claims column.
if stated_rhs is None:
return "unexplained", None
if abs(assets - stated_rhs) / assets > TOLERANCE:
return "genuinely broken filing", None
return "missing tag", NoneEvery branch returns a name. None returns an adjusted figure, and that is the property worth keeping if you write your own.
4. Rounding
Filers report in thousands or millions, and the rounding happens before the totals are struck rather than after. A balance sheet a few hundred billion dollars deep can show a one-unit difference between its two sides and be entirely correct.
Detection. The gap is small against total assets and nothing else accounts for it: no noncontrolling interest line, no mezzanine line, no shortfall against the filer's own stated total.
Pass it, and record the tolerance that let it pass. The second half is the part worth arguing about. A tolerance has to be a fraction of assets rather than a number of dollars: one calibrated on a mid-cap rejects every large bank, and one calibrated on a large bank waves through a genuine error at a small company. Half a percent of assets absorbs presentation rounding at any size.
I am not naming a current member. The category holds a handful of filings and turns over with every ingest, so a ticker printed here would be stale before you checked it.
5. A filing that genuinely does not balance
The company made an error. It should be rare for an audited public filing, and the interesting part is proving it rather than assuming.
Detection is one field.
LiabilitiesAndStockholdersEquity is the filer's own stated
right-hand side, the total printed at the foot of the claims column. Where they
publish it, two questions are available instead of one.
Does the filing balance against its own stated total? Assets against the stated right-hand side. If those two disagree, the filer's arithmetic disagrees with itself, and no extractor repairs that.
Did you recover everything the filing put there? The stated total against your reconstructed liabilities plus equity. If the filing balances against itself and your sum falls short, the shortfall is a line you did not read: category one, and yours.
That single field separates a missing tag from a broken filing, and without it neither claim is available. Where a filer publishes no stated total the honest label is unexplained, because assigning it to whichever bucket reads better is how a classification stops being a measurement. Business development companies and commodity trusts land here most often.
Flag the filing. Say what is wrong. Do not adjust.
Then the honest part. Across three separate measurements of every testable filing I hold, the count of genuinely broken filings has been zero each time. Zero is the right answer for audited public companies, and worth stating precisely because it is the outcome most worth measuring: every gap that looked like a filer's error turned out to be a term I was not reading. So I have no real example to show you, and inventing one would undo the point of the category.
What the classification is worth
It turns a percentage into a list.
"Ninety-nine percent of filings reconcile" tells you nothing you can act on. You do not know whether the remainder is a kind of company you hold, or whether those failures are the provider's extraction or the filers' arithmetic. The number asks to be believed rather than read.
"Three hundred filings failed, here are the tickers, and here is the reason for each" tells you which of your positions are affected and which of the failures deserve any attention at all. A rounding flag on a company you hold is nothing. A missing-tag flag on the same company means a component of that balance sheet is absent from your data, and you know to go and read the filing yourself.
Two questions to put to a data provider follow directly. Which filings failed? And why did each one fail?
Without the first, the output cannot be audited at all. Every figure arrives carrying the same implied confidence, and the wrong ones look exactly like the right ones. Bad financial data shows up not as a crash but as a plausible number from a real filing, which is the argument in the companion note on duplicate tags.
Without the second, the extraction cannot be fixed either. A failure with no category is a filing somebody inspects by hand once and never again. A failure labelled "mezzanine" is a mapping to write, and writing it closes several hundred filings instead of one. A bucket that fills up is a specification.
How this runs on BalanceProof
Every filing is reconciled before it is stored. The plain identity first; where that does not close, the noncontrolling interest and the mezzanine block are tried singly and together, and whichever basis closes with the smallest remaining gap is the one recorded. The company page names that basis in words, because this balances once the minority interest is included is a different statement from this balances.
Anything that does not close keeps a category: missing tag, rounding, genuinely broken, or unexplained where there is no stated total to referee with. The methodology page carries the count for each, read off the database when the page renders rather than written down once. There is no accuracy rate on it and there is not going to be: the denominator moves every time coverage improves, so a number that falls as the data gets better is not measuring the data.
Figures are served as reported. A filing that does not balance is drawn with the gap shown and the reason named rather than adjusted until the columns agree. JPM is the worked example worth opening first: twenty-three Assets tags, one consolidated set among them, an identity that closes. Looking a company up needs no account, and pricing covers the API and the bulk dataset.
If you find a sixth
Run the classification over your own data. It needs three totals per filing, the filer's stated right-hand side where published, and the two extra terms. The function above is the whole of it.
Then look at what does not fit. A filing whose failure is none of these five is the interesting result, and worth more to me than another confirmation of the categories I have. These five are the residue of what has broken so far, which is exactly the kind of list that stays incomplete without anybody noticing. Send me the ticker and the period and I will look.