# AIamond source record — method v2 Registered: 2026-11-03. Operator: Lennart Schreiber, aiamond.com (imprint: https://aiamond.com/imprint). This file is published unchanged at https://aiamond.com/method/source-record-v2.md. Its SHA-256 is printed on https://aiamond.com/how-reliable-is and in the change log (https://aiamond.com/changelog). Any change to this method is a new file with a new version number; this one is never edited. ## What changed from v1 Version 1 (https://aiamond.com/method/source-record-v1.md) computed chance from the shares of a whole cell, all markets together. Its first production run (2026-11-03) showed why that is wrong: sources write differently about different markets (more neutral about currency pairs, for example), and markets move differently (currency pairs stay inside the sideways band more often), so a cell that mixes markets rewards the mix itself. The headline came out "better than chance" at 1 and 7 days (+3.5 and +2.3 points); with chance taken within each market, a check on the same rows the same day gave about one point or less at every horizon. Version 2 takes chance within each market. Nothing else changes. The v1 run is not a result and is not published as one. ## Question How often did the market move the way a news source's article pointed, measured from the first price a reader of the article could have traded, and is that more often than chance for the same mix of calls? This answers "did the market move that way?" (a descriptive question) from a tradable start. It is a hit rate, not a return: it says nothing about the size of a move, costs or profit, so it does not answer "could a reader gain by acting?". AIamond makes no forecasts of its own. ## Data - One **check** is one article-based report and one horizon. Horizons: 1, 3 and 7 trading days (for crypto and currencies 1, 3 and 7 periods of 24 hours). - A language model reads the article and labels the market it is about and its tone from −1 to +1. Tone above +0.25 is a call "up", below −0.25 a call "down", anything in between "sideways". - The price move is measured from the first tradable price after the article's publication time (`start_kind = tradable`). Rows still measured from the previous close (`legacy`) are excluded. - The move counts as "up" or "down" when it exceeds 0.2 % × √days, otherwise "sideways". A check is a **hit** when the call matches the move. - Only sources with an admitted rights record are included: rights status `admitted`, `permitted` or `grandfathered`, and a publisher kind (official body, news agency, editorial news, specialist press, crypto editorial; kind unknown only when grandfathered). Aggregators, press-release wires, recommendation services and user-posts platforms are excluded, and so are sources whose content was deleted. ## Statistics per cell A cell is a group of checks: one source with all horizons pooled, one source and one horizon, one bundle of sources (by kind or by language) and one horizon, or all sources and one horizon (the headline). - **Hit rate** = hits / checks. - **Chance** is taken within each market (one price symbol), each check against the other checks of the same market in the cell. For market m with n(m) ≥ 2 checks and H(m) hits, let S(k) and W(k) count its checks that said and went way k (up, down, sideways); its expected hits are [S(up)·W(up) + S(down)·W(down) + S(sideways)·W(sideways) − H(m)] / (n(m) − 1). A market with a single check in the cell counts at its own result. Chance of the cell = Σ expected hits(m) / Σ n(m). It is the rate the same calls would reach against the moves of the other checks in their market, so knowing which markets move little or a lot earns nothing, and a check is never compared with its own move. - **Edge** = hit rate − chance, in percentage points. - **Interval**: checks close in time share market moves (the same day, and for longer horizons overlapping price windows), so they are not independent. The resampling unit is therefore a block of calendar days (UTC, by article publication time), counted from Monday 2020-01-06: 7 days for the 1- and 3-day horizons, 14 days for the 7-day horizon and for all horizons pooled. 2,000 bootstrap draws of blocks with replacement (fixed seed 20261103) give the standard error of the edge; the 95 % interval is edge ± 1.96 × that standard error. - **p-value**: two-sided, from z = edge / standard error under the normal approximation. A standard error of zero gives p = 1 (no verdict other than "not distinguishable"). - **Detection limit**: 2.80 × standard error, the smallest edge this cell could show with 80 % power at the 5 % level (before correction). ## Which cells get a rate, and the verdict - A cell with fewer than **300 checks**, or whose checks fall into fewer than 5 blocks, gets the verdict "too few data" and no rate is published for it; its number of checks is still shown. - No published number may reveal a withheld rate by subtraction. So a source's three horizon cells carry rates only when all three have a rate (otherwise only its pooled cell can), and the headline and the bundles count only sources whose three horizon cells all carry rates. - **Multiple testing**: within each family the p-values of the cells with a rate are corrected with the Holm method. Families: (A) sources with all horizons pooled, (B) sources × horizon, (C) bundles × horizon, (D) the headline, all sources × horizon. - Verdict: **better than chance** when the Holm-adjusted p-value is below 0.05 and the edge is positive; **worse than chance** when it is below 0.05 and the edge is negative; otherwise **not distinguishable from chance**. - The headline is family D. ## Limits Daily and hourly bars, one symbol per market. A market with a single check in a cell adds nothing to the edge. The model's labels agree with human readers in about 75–80 % of cases (a guess, not measured on this data); older reports were not all forward-looking (59 % in a sample of 800). Blocks do not remove all dependence: checks at the end of one block and the start of the next can still share moves. A source that is "not distinguishable from chance" may still be better or worse than chance by less than its detection limit. ## Corrections and right of reply Errors and replies from a named source: info@aiamond.com. Corrections are listed in the change log.