Change log: what changed in our numbers
AIamond publishes how often news sources were right and how that was measured. When we change how a number is measured, or find a mistake in the data behind it, it is recorded here with the date, what was wrong and what it does to the published figures. Nothing is removed from this list.
- Measurement methodCorrection running
New method for the source table: intervals, correction for many sources, tradable start only
From 3 November 2026 the source table at /how-reliable-is follows a written method, registered before its numbers were published: /method/source-record-v1.md, SHA-256 868f91c926ab2f151262048a9e9fb97d929a6b8e9882e3dfab20ffafbb234832. What changes: (1) only checks from the first price a reader could have traded count, not those from the previous close; (2) only sources whose rights record admits them count; (3) every difference to chance comes with a 95 % interval, in which checks close in time are resampled together in blocks of 7 or 14 days because they share market moves; (4) a source is called better or worse than chance only if that holds after correcting for the number of sources tested (Holm method), no longer at a fixed gap of 5 points; (5) no rate is shown for fewer than 300 checks, and no published total lets such a rate be worked out by subtraction, so the overall figures count only sources that have a rate at 1, 3 and 7 days; (6) the same numbers appear for groups of sources by kind of publisher and by language. The source pages and the comparisons no longer show hit rates per market or per day, and comparisons follow each source's verdict against chance instead of the gap between two raw rates. These numbers answer "did the market move the way the news pointed?", not "could a reader gain by acting?". The figures of the first run will be added to this entry.
- DataDone
TradingView user ideas deleted
On 3 November 2026 we deleted what we had stored from TradingView Ideas, a platform on which users publish their own chart ideas and trade tips. The deletion ran from 10:18 to 10:19 UTC and left no TradingView row behind: 63,485 reports (45 % of all reports), the 144,047 price checks made on them, 121,972 articles, 167,180 translations, 515 TradingView headlines in 103 Morning Briefs and 8 cached analyses. User posts with trade tips are not news from a publisher, and our feed rules do not admit them; the source had been switched off since 2 November 2026. What remains: the record that it was dropped, a log of how many rows were deleted, and texts we wrote ourselves before the deletion, such as the News Picture, which can still name TradingView among the sources it read. Backups made before the deletion still contain the deleted rows. TradingView made up about half of all price checks, so the combined figures changed. Measured right after, across the published sources (every source and time window with at least 300 checks, weighted by the number of checks): 130,253 checks, hit rate 35.92 % against 35.25 % by chance, a difference of +0.67 points; counting only checks from the first tradable price, 59,048 checks, 37.17 % against 36.61 %, +0.56 points. Both answer “did the market move the way the news pointed?”, not “could a reader gain by acting?”, and both lie far inside the ±5 points within which we call a result chance. They still move while older checks are re-computed from the first tradable price. The per-source tables have not listed TradingView since it was switched off.
- DataDone
Fewer sources: feeds without a clear right to be read are switched off
Since 2 November 2026 we have switched off 56 of the 90 news sources that were active: 19 feeds put together by third parties instead of the publisher itself (Google News, Yahoo Finance and similar); 3 Investing.com feeds (the other 9 were already off); 6 platforms for users’ chart ideas and trade tips, press-release wires and stock-pick services; 5 crypto outlets that do not meet our stricter rules for crypto media; 11 further feeds that our review rejected; and 12 feeds of ten publishers whose terms of use, as we read them, restrict automatic or commercial use, until each publisher allows it in writing. The other 34 stay active: 24 admitted after our review and 10 kept on probation until 1 February 2027 and decided again then. Every source now carries a record of its type and of where its right to be read stands (admitted after our review, allowed by the publisher, on probation, unclear or dropped), and the database no longer lets a source be switched on without a record that allows it. Switching a source off deletes nothing: the articles, reports and price checks already stored stay stored. The per-source tables and the source pages list only active sources, so the switched-off ones are gone from them. The combined figures will be recomputed on the sources that remain and will change; each change is recorded here.
- DataDone
One story, one article: repeats removed
Until 1 November 2026 the same news page was often stored several times: when a site rewrote a headline, or a Google News summary listed other related stories, our system took it for a new article and analysed, checked and counted it again. Of 1,097,208 stored articles on 1 November 2026, 238,055 (21.7 %) were such repeats. They were removed in the background, together with the 30,056 reports and 64,965 price checks that belonged to them; 1,296 reports were kept by moving them to the remaining copy. The run finished on 1 November 2026 at 19:04 UTC. Since then no two stored articles share a page address and a database rule makes a repeat impossible; new articles are compared by the address of the page, so a rewritten headline is no longer stored again. Every story now counts once in every hit rate, count and sentiment figure. Measured on the live data after the run, across the published sources (every source and time window with at least 300 checks, weighted by the number of checks): 254,672 checks, hit rate 35.30 %, edge over chance −0.11 points. These figures still move while older checks are re-computed from the first tradable price. Google News Business, the most repeated source, went from 20,539 to 10,836 checks, and two sources (BBC Business, NOS Economie) now have fewer than 300 checks and show “too few”. In a test run on a snapshot before the clean-up, no source with at least 300 checks changed its verdict.
- DataDone
Publisher summaries are shown cut to 150 characters
Until 1 November 2026 the summary text that publishers send in their feeds appeared in full on our site, in our own RSS feed and in our Discord and Telegram posts. Those texts belong to the publishers. Now every public place shows at most the first 150 characters of a summary, followed by “[...]”; the headline always links to the original article. The full text stays stored internally for the analysis. Publishers may ask for less or allow more; any change will be noted here.
- DataCorrection running
Currency pairs were checked against the wrong market
Until 1 November 2026 a text rule filed currency pairs such as EUR/JPY, EUR/GBP and NZD/JPY under “Japanese stocks”, “Australian stocks” and “UK markets”, so their price checks ran against the Nikkei, the ASX and the FTSE instead of the currency pair. Of the 5,533 stored reports under those three labels, 1,617 are about a currency pair (1,174 “Japanese stocks”, 206 “UK markets”, 237 “Australian stocks”). New reports are filed correctly since 1 November 2026. The older ones are being moved to their currency pair and re-evaluated in the background; the old label is kept with each moved report. The rule only moves clear cases. We estimate that 15 to 20 % of the currency reports under these labels, mostly yen stories without a pair code in the title, are not recognised and stay where they are, so the figures for Japan, Australia and the UK remain less reliable than the others. On the old method, taking the moved reports out changes the share of correct calls of these three markets by 1 to 3 points. Currency pairs that were never evaluated before, such as AUD/USD and USD/CAD, are evaluated now, so the forex figures gain rows. The figures after the re-evaluation will be added here.
- Measurement methodDone
Checks now start at the first price a reader could trade
Until 1 November 2026 the check “did the market move the way the article said?” started at the close of the day the report was written. Nobody can trade at a price that is already behind them. Now the check starts at the first tradable price after the article was published: the session open if the article appears before the open or after the close, the next hourly price during the session. The difference is large. The same 23,469 fresh reports about stock indices (analysed within two hours of publication) show an edge over chance at 1, 3 and 7 days of +7.2 / +2.6 / +2.0 points when measured from the previous close, +3.5 / +0.4 / +0.7 from the session open and +0.8 / −0.5 / 0.0 (± 0.3) from the first tradable price. Against the previous close, news looks right mostly because it describes moves that had already happened. Older reports are re-computed in the background; until then the source pages mix old and new rows.
- DataDone
New reports carry a word-for-word quote from the article
To make “the source says …” checkable, the analysis now has to copy one passage from the article text word for word, and our code checks that it really occurs in the text. A report keeps its direction only if the quote is found and the text contains a real outlook, not a recap; otherwise the report keeps its market but has no direction. In a test on 200 articles the model's quote was found in the text in 139 of 142 cases; the 3 invented ones would have been dropped. Reports written before 1 November 2026 have no quote. For them, an independent re-reading of 800 random reports found that only 59 % of those carrying a direction were truly forward-looking.
- DataDone
Late reports now carry the time of the article
Articles that were analysed days after they appeared were given “now” as the time of their report, so feeds, buzz and the checks treated old news as new. In October 53 % of the reports were written more than six hours after publication. New reports now carry the publication time of the article; the time of the analysis is stored separately.