The filing-change monitor

The test, and what it finds

Each pair is a bank's filing against its own filing one year earlier (10-K vs prior 10-K, 10-Q vs the same fiscal quarter's 10-Q, CMN's pairing). Change intensity is 1 minus similarity. Pairs are sorted into quintiles within form and filing year, so time and form effects wash out, and the table reports the median change over the four call-report quarters after the filing: noncurrent loans as a share of gross loans (FDIC NCLNLSR), the common equity tier 1 ratio (FDIC RBCT1CER), and this site's 0-100 composite bank score. 7,355 of 8,613 pairs enter the test (the rest lack an FDIC cert match or a complete forward window).

The same sort on Jaccard similarity (distinct-term overlap rather than frequency-weighted overlap):

The finding is a null, and it is reported as one. No column shows a monotone gradient across quintiles. The biggest changers (Q5) show a median four-quarter noncurrent-loan move of -0.01 pp against -0.02 pp for the least changed (Q1), a gap of +0.01 pp: economically nothing. The composite-score column leans, if anything, the OTHER way (Q5 +0.3 vs Q1 -0.1, a gap of +0.4 points), and not monotonically. On this panel, over this horizon, filing-change intensity does not predict call-report deterioration. That does not contradict CMN, whose outcome was equity returns on a broad panel, and it is itself informative: whatever bank 10-K risk sections churn on, it is not a one-year-ahead signal of their own balance-sheet decay.

The SVB check

Reported whatever it shows: SVB Financial Group's last 10-K pair (fiscal 2022 against fiscal 2021, filed 2023-02-24, fourteen days before the March 10, 2023 failure) has cosine similarity 0.982 and Jaccard similarity 0.660. That places it at percentile 70.4 of change intensity among all 2,277 10-K pairs, and in quintile 4 of its filing year's 10-K change sort (all pairs, not only the outcome-testable ones): elevated, not extreme. Its sharpest rewrite came earlier: the fiscal 2019 pair sits at percentile 97.3.

For context, SVB's earlier 10-K pairs: fiscal 2016 cosine 0.994, fiscal 2017 cosine 0.994, fiscal 2018 cosine 0.994, fiscal 2019 cosine 0.900, fiscal 2020 cosine 0.977, fiscal 2021 cosine 0.978.

How similar are consecutive filings?

CMN report that most filings change very little year over year, which is what makes the changers informative. The bank corpus reproduces that: the median 10-K pair has cosine similarity 0.987 (10th percentile 0.961, 90th 0.995), and the median 10-Q pair 0.954. Jaccard runs lower by construction (0.682 median for 10-Ks, 0.513 for 10-Qs): it ignores term frequency, and the long tail of once-used terms overlaps less than the frequency-weighted core vocabulary.

The returns leg, and why it is not here

CMN's own outcome is equity returns. The estate's daily log-return panel covers 30 large financials from 2020, and only 21 of the 314 corpus banks are in it: too few for a quintile sort anyone should trust. The returns leg is therefore reported as exactly that overlap count and skipped, rather than run on a sample that cannot support it.

Method, verbatim

• Corpus. The same EDGAR corpus as the language page: primary 10-K and 10-Q documents (amendments excluded), filed 2016 onward, for the fetched SIC 6021/6022 slice of 314 banks. See the language page for the corpus boundary; 297 of those banks have at least one same-form year-over-year pair.
• Pairing. CMN's rule: a 10-K pairs with the prior year's 10-K, a 10-Q with the SAME fiscal quarter's 10-Q one year earlier. Implemented as: same bank, same form, prior document whose period of report is 300-430 days earlier, nearest to 365 days; one document per (bank, form, period), larger kept.
• Similarity. The document's HTML with script/style blocks, markup and entities removed, whitespace collapsed, lowercased; tokens are maximal [a-z0-9] runs. Cosine similarity over unigram+bigram term-frequency vectors of the FULL text, and Jaccard similarity over the distinct-gram sets: CMN's two main measures. No stopword list, no stemming, no model. Change intensity = 1 - similarity.
• FDIC join. EDGAR registrants are holding companies; call reports belong to bank charters. The crosswalk matches the registrant's current and former EDGAR names against FDIC BankFind holding-company (NAMEHCR) and institution (NAME) names after suffix and abbreviation normalization, preferring the active, largest institution: 286 of 297 pair-universe banks matched. Unmatched banks keep their similarity rows and drop out of the outcomes test only.
• Outcomes. Baseline is the last call-report quarter strictly before the filing date; the change is measured to the quarter one year later. Noncurrent rate is FDIC NCLNLSR ("N/C LNS & LS/GROSS LNS & LS"), capital is FDIC RBCT1CER ("COMMON EQUITY TIER 1 RATIO"), and the composite is this site's peer-scored 0-100 bank score. Pairs whose forward quarter is not yet published drop out, so 2026 filings carry no outcomes yet.
• What this page does NOT claim. This is not CMN's test: their outcome is returns on a broad CRSP panel; this is call-report deterioration on 297 banks, of which 7,355 pairs are testable. Three outcomes on two measures invite multiple-comparisons doubt, medians of a mostly healthy panel are small numbers, and a null result here is a result, not a failure. No causal claim is made in either direction: a bank may rewrite its filing because conditions already turned.

Sources: SEC EDGAR 10-K/10-Q filings (US government work, public domain); FDIC BankFind institutions and financials (public domain). Reference: Cohen, L., C. Malloy and Q. Nguyen, "Lazy Prices", Journal of Finance 75(3), 2020, 1371-1415. Companion pages: the language of bank risk · as-filed XBRL · bank scores.