How the numbers are made

Every bookmaker builds a margin into its odds. Remove that margin from a sharp bookmaker’s price and you get a good estimate of the real chance of each outcome — a fair price. A pick is a price, somewhere in the market, that pays more than that fair price. Each pick on /picks carries one of three labels, saying where its fair price came from:

• Sharp line — the fair price comes from the sharpest bookmaker, the one whose closing odds are the best-documented predictor of results in the betting research. We publish a pick when another bookmaker pays at least 3% more than that fair price, at odds up to 4.00. No prediction model is involved.
• Consensus — the same idea where no single line is sharp enough to trust: the fair price comes from several bookmakers agreeing, margins removed. The label says how many.
• Model — the fair price comes from our own model. It combines a goal-rate model with a machine-learning classifier (form, team ratings, expected goals, line-ups, injuries, how the market has moved), and it leans on the sharpest bookmaker’s price where that market is sharpest. It is retrained every week on finished matches; a new version replaces the old one only after we have checked it on matches it was not trained on.

The two kinds of edge are measured differently and cannot be compared. A sharp or consensus edge is an expected return (+3% means 3% above break-even). A model edge is the gap between our probability and the bookmaker’s, in percentage points — at long odds, a model edge of 16 points is a far bigger claim than a 3% sharp edge, not a five-times-better pick.

Each pick also shows a break-even price: the lowest odds at which it is still worth taking. Odds move after a pick is posted, so check the price you are actually offered against it.

A strategy is one rule for finding picks (a market, a set of leagues, an edge threshold). We have tested well over a hundred and retired most. Each one has a status, shown next to its picks:

• Experimental — visible only to us. Nothing is published.
• Testing — its picks are not published: only Active strategies post to /picks and our Telegram channel. A Testing strategy keeps its own record on /performance, but it does not count in the headline. It is promoted to Active after 50 settled picks that beat the sharp closing line on average (see below).
• Active — its picks go to /picks and the Telegram channel (on Telegram, at most one pick per market per match), and they count in the headline record.
• VIP — a small number of strategies post only to our paid VIP channel. Their picks are sealed before kick-off like every other, and each has its own record, shown once the picks have settled. They are not in the headline. Counted at the best price we found when we posted it. Prices move; check yours before you bet. VIP details
• Retired — strategies we stopped. Nothing is deleted: their records stay on /performance under “retired strategies”, losers included.

A pick counts in the record if its strategy was Active when the pick was made — and it stays counted for good. Pausing or retiring a strategy never removes its picks; it only stops new ones from counting. Picks a strategy made while it was Testing never count, even after it is promoted.

We show two numbers: the all-time record (every such pick, including strategies we have since paused or retired) and the current line-up (the same rule, over the strategies Active today).

Three more rules. Some strategies have an edge ceiling above which we never show or send their picks (their highest-edge picks tested worse); a pick above it is left out of the headline, because no reader could have followed it, but it still counts on the strategy’s own row. And when we change a strategy’s rule we do not restart its record: a strategy whose record spans several rule versions says so on /performance. And a pick is graded at the best price available when it was made — unless that price was more than 1.25× the sharpest bookmaker’s price and no other bookmaker came within 5% of it; then it is graded at the best price another bookmaker actually offered (a reader could probably not have taken the first). We only judge this when at least three other independent bookmakers were still quoting; otherwise the pick is counted and marked as not price-checkable. Since 4 October 2026 a bookmaker quoting the price at our next price snapshot, up to 35 minutes after the pick, also counts as confirming it — our snapshots run every 30 minutes, while picks read live prices in between. The pick's own bookmaker never confirms itself, and on Asian handicap we count two bookmakers that quote the same price most of the time as one source. Each pick stores which version of this check judged it; picks judged before that date keep their verdict.

Every change to these rules is dated, with its effect on the figures, on the corrections page.

A few hundred results are mostly luck. A better early signal is whether our price beat the market’s final price before kick-off — the closing line. Sharp markets move toward the right price as a match approaches, so a bettor who keeps getting better odds than the close is usually right about value, whatever the next ten results say. That gap is closing line value (CLV).

Which CLV comes first. The CLV we show first, on /picks and /performance, is against the sharp closing line: the sharpest bookmaker’s closing price with its margin removed, or a consensus of five or more bookmakers where it has none. /picks also shows a second figure, labelled “vs book close”: the price against that bookmaker’s own closing price, margin included. It reads low by design for picks that exist because a bookmaker was slow to move, so it is a check, not the verdict.

A closing price only counts if it is fresh. If the sharpest bookmaker’s last price had not moved for more than two hours while other bookmakers or a betting exchange clearly had, we grade the pick against the exchange’s close or an adjusted price instead — or give it no sharp closing value at all.

Every figure is computed at €10 flat per pick at the price available when the pick was made: a win pays €10 × (odds − 1), a loss costs €10. One fixed unit keeps the record reproducible — anyone can re-derive it from the public ledger with no bankroll assumptions — and it is the unit we settle rival tipsters’ picks at too.

The pages then let you view the same record at the stake you would bet, from €5 to €250 per pick; they open at €25, which is why /performance and /performance/details show €25 figures. Changing the stake only rescales the money figures (profit, drawdown). ROI, CLV, hit rate and pick counts are the same at every stake. See it on the full record.

For each public tipster or model we compare with, we collect its published pick history and keep only what both sides cover: the same dates, the same markets, settled picks only. Both sides are then settled at the same flat stake. Everything else is dropped from both sides, so a comparison row counts fewer picks than our headline and can sit above or below it — both are real, they measure different sets.

Where a tipster’s own figures use prices no bookmaker we track ever offered, we show its claim next to its picks re-settled at the best real closing price. Every matched pick on both sides is published in the public ledger, so anyone can recompute both numbers. Our reviews go through individual tipsters in detail.

Four pieces of evidence, each proving something different:

• Before kick-off: the seal. Every few minutes each batch of new picks is turned into a short code that changes if a single character of a pick changes, and that code is signed by an independent timestamp authority (DigiCert, with its own clock). After kick-off the pick and its signed seal are published, and anyone can check them on the verify page — the proof covers every pick, posted or not. It started on 28 September 2026.
• Before kick-off: the Telegram channel. Many public picks are also posted to our Telegram channel before the match. Telegram’s post time is one we cannot edit.
• After the matches: the nightly ledger. Every night a snapshot of every settled pick is committed to the public ledger on GitHub. Each snapshot holds the full record, so comparing two nights shows that no earlier result changed. It is written after the matches, so on its own it does not prove when a pick was made.
• Unchanged since that night: the Bitcoin timestamp. Each nightly snapshot is also timestamped on the Bitcoin blockchain with OpenTimestamps, which proves the file has not been altered since that night without trusting GitHub or us. It proves nothing about when the picks inside it were made.

The nightly snapshots run 24 June – 12 July 2026 (8 July missing); there are none from 13 July to 20 September 2026; since 21 September there is one every night. Picks settled in the gap are in every snapshot from 21 September, but their stamp only proves they have not changed since then. Removed picks, with the reason, are listed on the corrections page. The settled record is also available as JSON at /api/v1/track-record.

Our model and software are private. None of the checks above needs them: the ledger, the seals and the JSON are enough to recompute every published figure.