This page describes how the Accuracy Index is calculated. It does not report results. Each edition’s figures are published in its own study, listed under Published editions below.
What the Accuracy Index measures
The Accuracy Index is the share of prediction services that finish a fixed period in profit. A service counts as profitable if the combined result of all its scored predictions, at a flat stake, is above zero.
The index covers football only, and one market only: the match result (home win, draw or away win). Each edition covers one fixed calendar period, stated in its study.
Why we built it
We wanted one measure, applied on identical terms to every service we track, of how many prediction services make money over a fixed period. It is a finding about the industry, not a verdict on any single business.
What we track
For each prediction we record the match, the outcome picked, the service and author, the market price and the settled result.
The market price is the price Tips.GG records for each match. Every prediction on a match is settled at that same recorded price, so no service gains an advantage from the price it was given. We check this for every edition: no match may carry more than one recorded price.
These prices include a margin. We keep it in because betting prices always carry one: a service must overcome it to finish in profit. Each study reports the average margin for its period, calculated from the same recorded prices used to settle predictions.
We publish the formula, the sample and the results table. We do not publish the raw tracking data.
How a single prediction is scored
Every prediction is treated as a flat $10 stake. Real stakes are unknown, so a fixed stake is the fair basis for comparison. The unit does not affect the result: ROI in percent is identical at $1 or $10.
Win: profit = $10 × (price − 1).
Loss: profit = −$10.
A service’s ROI is its total profit divided by (number of scored predictions × $10), expressed as a percentage. A service is profitable if its total profit is above zero.
Only predictions with a settled win or loss are scored. Predictions on cancelled matches, or matches whose data was lost, are removed before calculation.
Predictions on played but not yet settled matches are also excluded. To keep these few, we finalise each edition no earlier than one to two weeks after its period ends.
We also report accuracy: the share of scored predictions that won. Accuracy is not the index. A service can win most of its picks and still lose money if the prices are short.
Which sources are included
A website is included if it published at least 50 predictions in the period. A channel on a platform needs 200. Smaller outlets are left out.
The thresholds differ because the same number means different things. In the first edition’s data the median website published 1,126 predictions over eight months; the median channel published 221. Fifty predictions from a website is a nearly dormant operation. Fifty from a single person posting picks is an ordinary couple of months.
That difference shows up directly in the results. Among channels with 50 to 199 predictions, 53% finished in profit; among channels above 200, 26% did — the same range as websites. The gap is the length of the run, not skill: over 50 bets one good fortnight shows as a profit that disappears over 500. Left at 50, the top of the table would have been an account with exactly 50 picks, in the table only because it had just cleared the bar.
Neither threshold decides the headline figure. Moving the channel threshold from 100 to 500 moves the index between 25.7% and 21.0%; moving the website threshold from 30 to 100 moved it by about three points. The full figures appear in each study.
| Minimum predictions for a channel | Accuracy Index, first edition |
|---|---|
| 100 | 25.7% |
| 200 (used) | 23.0% |
| 300 | 21.6% |
| 500 | 21.0% |
What counts as a source
The unit we measure is a single outlet: the thing a reader actually follows. On the web that is a website, and every author publishing under it counts towards its result. On a platform it is a single channel or account, not the platform itself.
A YouTube channel is no less a separate outlet than a website: it has its own author, its own audience and its own record. Counting all of YouTube as one row would be like counting the entire web as one row, while we count websites one by one. So each platform is split into the channels publishing under it, and those that clear the channel threshold appear as their own rows.
A channel that republishes predictions already counted elsewhere in the table is excluded, because the same predictions would otherwise enter the index twice. In the first edition that rule excluded our own Telegram channel, which reposts picks already counted under tips.gg.
A few rows are community-voting feeds: an aggregate of what a platform’s users picked rather than one author’s calls. They stay in the index — a reader can follow them exactly like any other signal — and each study says which rows they are.
When two names are the same publisher
Labels shown to be the same publisher are merged into a single row before calculation. Counting one publisher twice would distort both the index and the leaderboard: one operator publishing under several names would appear as several independent services, and could hold more than one place at the top of the table.
Any one of three findings is enough to merge:
- the same author accounts publish under both labels;
- the labels make near-identical picks on the same matches: above 95% agreement across at least several hundred shared matches;
- the same domain is recorded for both labels.
The check covers the whole dataset, not a visual sample. A merged row combines the predictions recorded under each of its names. Where the same pick on the same match appears under more than one of those names, it is counted once. Accuracy and ROI are then calculated from the combined predictions, not averaged from the separate figures.
Merging can change the leaderboard. A service that ranks highly on its own can drop out of the top places, or out of profit, once it is combined with the other names of the same publisher.
The check works in both directions. Similar names, a shared brand style or partly overlapping picks are not enough on their own. A pair is merged only if one of the three conditions is met. Each study lists the merges applied and the pairs checked and kept separate.
How services are named
Every outlet that clears its threshold is named, whatever its result, and so is Tips.GG. A table that withheld the losing names would ask a reader to take most of the finding on trust, and the result of one period is a measurement, not a verdict on a business — each study says so in as many words.
Where a service is named, we use its domain rather than a styled brand name. Each prediction we record carries the address of the page it came from, so the domain is part of our data and can be verified. A styled brand name would be our own invention. Where a service publishes across more than one domain, the row carries the domain that accounts for the majority of its predictions; where the split is close to even between zones of the same brand, the row carries the brand the domains share, without a zone. Each study flags every row where a single domain did not clearly dominate. Where a domain cannot be established from the data, the row carries the identifier the service is recorded under.
A merged row is published under the group’s primary brand domain: the root or best-known domain, not a variant. Where no clear primary exists, the domains are shown separated by slashes.
Channel rows carry the channel’s own name with its platform in brackets, for example VoVo (YouTube). We use the name as our tracker records it rather than inventing a display name.
We include ourselves
Tips.GG is included in every edition on the same terms as every other service: same period, threshold, prices and stake. It is always named, whatever its result. The Tips.GG figure reflects the site’s aggregated output, built from our editorial team’s predictions.
What the index does not tell you
It covers football match-result predictions only. Esports, other sports and other bet types are not included.
The stake is notional. Real stakes are unknown.
Each match is settled at one recorded price. That is not necessarily the price any individual reader could have obtained.
Prediction timestamps in our records are not reliable enough to state exactly when a tip became public. A small share of records carry a timestamp later than kick-off. Each study reports their win rate and whether excluding them changes the index.
Outlets below their threshold do not appear at all. On the channel side that leaves out most of a platform’s smaller accounts.
Merges rely on what our data shows. Two publishers linked only by ownership, without shared authors, near-identical picks or a shared domain, are still counted separately.
We do not test for survivorship bias. Services that stop publishing mid-period are rarely numerous enough in a single edition to support a conclusion in either direction.
Each edition describes a finished period. It does not predict future results.
The author leaderboard on our site uses different rules. It ranks individual authors, requires 12 predictions rather than 50, uses a rolling window and shows only authors in profit. Its figures are not comparable with the index.
Published editions
Edition 1: 1 January to 31 August 2026.
Using our data
Each study publishes its results table as an HTML table inside the study. The table lists, for each included outlet: its name, number of predictions, accuracy, ROI and period. Every figure a study publishes is checked against a row of its table.
The index and the tables are free to cite and reuse, with attribution to Tips.GG and a link to the original study.
Questions and corrections: [email protected].
You can view all TipsGG research here.