01The summary
The quarter went into one concrete job: publishing the dividend per share of IBEX 35 companies with the issuer's document behind every figure. Doing that meant reading the XBRL accounts filed by 311 European issuers and checking them against what each company actually announces. That uncovered errors that had been there for a while.
- A code defect dating from version 29 (July 2026) that let one report's comparative column overwrite another report's main year, across 311 issuers. Bankinter gave it away, in the draft of its page.
- 49 earnings-per-share rows in the wrong scale: pence as pounds, cents as euros and a stray factor of a thousand. Unicaja showed €95 where it earns €0.095.
- A whole series shifted by a year (Endesa) and a total that already contained the tranche being added to it (Acerinox), because the same pair of IFRS concepts means four different things depending on the issuer.
- Four tickers serving another company's accounts, matched by name similarity.
- Two figures the external check caught before publication: ArcelorMittal with a row shifted by a year and Sabadell with a 0.75 that no source supports.
- This week's: Mapfre with a news aggregator as the source of its 2025 figure, visible since 4 October, and four new pages (Telefónica, Aena, Logista and Indra) whose drafts talked about «three filings» that were not on the page.
Of all this, what was actually visible on the site was the earnings per share in the wrong scale, the four tickers serving other companies' accounts and Mapfre's source; the rest was caught in the draft of the pages before they went live on 4 October. Both lists, dated, are in the corrections register. This report tells what is behind those rows.
02The comparative column was overwriting the main year
Our own team found it reviewing the draft of the Bankinter page by hand: the 2023 dividend dropped to €0.1422 from €0.3117 in 2022 with higher earnings per share. Bankinter reports its 2023 dividend twice, with two different values: €0.47002291 in its 2023 report, as the main year, and €0.14220835 in the 2024 report, as a comparative. The figure coming out was the comparative: 30% of the real dividend.
Since version 29 the code had carried a comment saying that a report's main year takes precedence over the same year-end read as another report's comparative. It was not implemented. Nothing knew which year was each report's main year; reports were walked newest to oldest and the first writer won. It was not a dividend bug: it affected any line item an issuer restates, across all 311 issuers, and the dividend only made it visible because a dividend series has a shape a human recognises at a glance and a restated balance sheet does not.
The main year is now derived from the data itself: the most recent duration period-end in the report. Fixing it changed ArcelorMittal, already reviewed: 2023 went from 0.50 to US$0.44 and 2022 from 0.44 to 0.38, which are the correct values, and a 2020 at zero appeared, which is true and is information: the dividend was suspended.
03Earnings per share in the wrong scale
Unicaja showed EPS of €95 for 2022, where it earns €0.095. It is the same scale error we already knew from dividends, but in earnings per share, and that one did appear in the table. And the damage did not stay in the cell: the share count is derived from net income ÷ EPS, and the share count feeds the cross-check that validates every dividend on the page.
There were 49 such rows in the universe, by three routes: pence tagged as pounds (Admiral 216.6, Berkeley 417.8, 3i 539), cents as euros (BBVA 2018: 75) and a stray factor of a thousand (Unicaja). The first check we wrote broke Bankinter: it turned its 2020 EPS from 0.35 into 0.00035, because it compared against the derived share count, and an odd share count can come from either field. It was rewritten against the issuer's own EPS series: Unicaja runs 0.045 · 0.544 · 95 · 0.098 · 0.217 and the 95 gives itself away; Bankinter runs 0.35 · 0.61 · 0.62 · 0.94 · 1.06 · 1.49 and is left alone.
04The same IFRS concept, four meanings
Endesa was shifted by a whole year in the draft: the page was going to say €1.5854 for 2023, which is its 2022 dividend; 2023 was €1.00. Acerinox was going to show «0.31 + 0.62» for 2023, where the 0.62 already includes the 0.31. Each figure was correct as a fact and wrong as an answer.
The reason is that the same pair of filed concepts, «recognised» and «proposed», means different things by issuer: at Iberdrola they are the year's two tranches and add up; at Inditex the recognised figure is the previous year's total; at Endesa likewise, with nothing proposed; at Acerinox the proposed figure is the year's total with the tranche inside. It cannot be derived from the filing uniformly, so there is no honest way to publish «the dividend for the year» by reading XBRL. What can always be stated is «what the report declares», and that is what the pages now say, with the four examples at the top rather than in a footnote. The answer to «how much does it pay» comes from a hand-curated file checked against each company's own announcement, which is a different thing.
05Four tickers were serving another company's accounts
Searching for duplicated legal entity identifiers in the universe turned up five. One is legitimate: IAG trades in London and Madrid. The other four were the very failure our own code warns about: matching by name similarity paired different entities. 3i Infrastructure was serving 3i Group's accounts; abrdn Asia Focus, Aberdeen's; Foresight Environmental Infrastructure, Foresight Group's; ICG Enterprise Trust, Intermediate Capital Group's. Balance sheet, income statement and dividend of the bigger cousin, under its ticker.
We did not look up the correct identifier by eye, which would be the same mistake again: they were withdrawn, with a written record so nobody restores them without finding their own LEI in GLEIF. The universe went from 311 to 307 issuers.
06What the independent check found before publication
The 32 curated figures of the first batch of dividend pages went through an external check with a self-contained package: 21 confirmed, 7 corrected, 2 downgraded to «paid during the year» and 2 corrections rejected; 4 new figures we did not have also turned up. The finding that matters most: Iberdrola paid a final dividend of €0.409 and the filing records 0.404, because the filing records the minimum announced, not what ended up being paid. The two corrections were rejected for adding to the dividend things that are not dividend: Iberdrola's attendance premium for the general meeting (€0.005) and a Mapfre participation dividend paid out of reserves.
The second pass, on 2024 and 2025, gave 26 of 28 confirmed and two real errors, both in the year added in one go. ArcelorMittal had its row shifted by a year: its 0.55 is fiscal 2024 (note 11.4 of its 20-F labels it so) and press releases call it the «2025 dividend» because it was paid in 2025. Sabadell had 0.75, which no source supports: the dividend against 2025 is €0.265 and the extraordinary dividend from the TSB sale is not charged to the year. And the method finding: the URL cited was the previous year's page. A source being the issuer's does not prove the document supports the figure.
This very week, in the second batch of pages, the review caught Mapfre 2025 linking to a news aggregator with the primary-source flag set (visible since 4 October; it now links Mapfre's notice to the CNMV) and, in the draft, four new pages without a filing table saying their figures came from «three filings» that were not on the page. The check now rejects aggregators.
07What changes from here
- The dividend pages' check goes from 17 to 41 assertions, ranking included: every change is recomputed from the curated file and an aggregator source fails the build.
- Every figure with its document, and the document has to support the figure. The primary-source flag is no longer enough; the review reads the document.
- Two independent reviews before every publication, one of the code and one of the figures. In both batches of pages published so far, both found something. None of it was cosmetic.
- A declared gap is worth more than an attribution that does not hold. Repsol pays against earnings and against reserves at the same time and has no clean total per year: it is published as «paid during the year» and the page says so.
- This report repeats every quarter, with the same rule: name, figure, document and which check prevents it.
Frequently asked questions
Why publish your own errors?
Because a site that promises every figure with its source is only credible if it says when it gets one wrong and how it finds out. A quality report without names or figures is advertising. Every case in this report carries the issuer, the value that was being served, the correct one and the document it was checked against, and the corrections register keeps it current between reports.
How were these errors found?
By three routes: our own team's hand review (Bankinter, Unicaja), automatic checks that test accounting identities and scales on every series, and two independent reviews before each publication, one of the code and one of the figures, with a self-contained package so they cannot drift from the data.
What is a scale error in filed accounts?
The issuer tagging a figure in a unit other than the one it declares: pence as pounds, cents as euros, thousands as units. Unicaja filed earnings per share of €95 where it earns €0.095. It is corrected only when another figure in the same report confirms it, and if nothing confirms it the figure is not published.
Can you read a company's dividend from its XBRL accounts?
Not uniformly. The same pair of concepts, dividend recognised and proposed, means different things by issuer: sometimes the year's two tranches, sometimes the previous year's total, sometimes a total that already includes the recognised tranche. That is why the pages publish what the report declares and, separately, the dividend announced by the company, checked by hand against its own announcement.
The figures in this report come from the quality control of Celsmar's dividend pages and fundamentals between July and October 2026, and each was checked against the issuer's or the regulator's document. Companies are named because their accounts are public and the check can be reproduced by anyone. Next report: January 2027.
This is information and analysis, not financial advice, and it contains no recommendation to buy or sell. See the disclaimer.