How we check the data
Every term in the registry is read from the note's final pricing supplement filed with the SEC, with the filing's own words beside it. Numbers and dates are checked against those words in code, and a note is published only when every check passes.
How each value is checked
Code checks every critical number and date. The quote has to appear in the filing, and it has to state that value. Categorical fields, such as the structure and the kind of downside protection, are read by the model and audited blind.
What it covers
104,491 notes priced through 4 November 2026, filed by BMO Financial Group, Bank of America, Barclays, CIBC, Citigroup, Deutsche Bank, Goldman Sachs, HSBC Holdings plc, JPMorgan, Jefferies Financial Group Inc., Marex Group plc, Morgan Stanley, Nomura Holdings, Inc., Royal Bank of Canada, TD Bank Group, The Bank of Nova Scotia, UBS, Wells Fargo & Company.
Measured accuracy
- Confirmed errors in the registry
- under 0.051% of terms, at 95% confidence: 14 errors in 43,472 terms from 3,825 notes drawn at random from the whole registry, 8 audits pooled.
- Notes with a wrong term
- 14 of 3,825 under 0.58% of notes, at 95% confidence.
Release 20261105-1157-94aaea0d, built 5 November 2026: measured on this release: 2 errors in 15,510 terms (1,369 notes), under 0.041%.
All runs since 11 Oct 2026, including corrected versions
| Date | Notes | Terms compared | Confirmed errors | Not settled |
|---|---|---|---|---|
| 1 November 2026 | 505 | 5,738 | 0 | 2 |
| 26 October 2026 | 864 | 9,772 | 2 | 2 |
| 25 October 2026 | 516 | 5,888 | 2 | 2 |
| 22 October 2026 | 1,600 | 18,228 | 8 | 8 |
| 21 October 2026 | 194 | 2,195 | 1 | 1 |
| 18 October 2026 | 23 | 263 | 1 | 1 |
| 12 October 2026 | 23 | 269 | 0 | 0 |
| 11 October 2026 | 100 | 1,119 | 0 | 0 |
The figure counts only notes drawn at random from the whole registry, about 11.4 terms a note. Audits also take extra recent notes and notes released by our newer checks, to find problems early; those notes are left out of the figure, and so are audits aimed at one issuer or period.
The bound on terms assumes each term is an independent draw. Errors cluster by note, so the note-level figure is the safer reading: 14 wrong of 3,825, under 0.58%.
Another 16 disagreements could not be settled from the filing's words. Counted as errors, the bound on terms rises to 0.094%.
4 of the 14 errors are structure labels, the family a note is filed under, and not the value of a term.
91% of the sample comes from 4 audits of 2,000 notes, the latest on 1 November 2026; the Sunday audit is set to read 1,700 notes from 25 October 2026, 400 of them drawn from the whole registry.
Published values were last repaired after an audit finding on 2 November 2026. No audit has measured the registry since, so these figures describe it as it was before that repair.
The audit compares 15 terms: structure, whether it is a basket, underliers, coupon rate, coupon frequency, coupon barrier, call level, first call date, kind of downside protection, barrier or buffer level, upside participation, cap, term, price to public and the issuer's estimated value. First call date is the only date the audit compares.
An independent reader from a second vendor (OpenAI) read 1,999 published notes drawn at random on 1 November 2026 and agreed with the registry on 99.57% of 27,212 terms. Every disagreement is read against the filing.
An audit describes a random sample. It can't rule out rare errors, and a value both readings get wrong in the same way agrees, so no audit sees it. The figures are a measurement, not a guarantee for every note.