Methodology

The foundation is an Elo-family rating fitted over more than 10,000 senior men’s internationals. Every competitive match updates the two teams’ ratings; friendlies count for less, and margin of victory matters. The rating gap between two teams — adjusted for host advantage and neutral venues — sets the expected strength difference for a fixture.

A Dixon-Coles bivariate Poisson model turns that strength gap into expected goals for each side and a full scoreline distribution — the probability of every plausible final score. Summing the grid gives the three match probabilities (home / draw / away) that you see on every match page, plus the top-scorelines table. Tournament forecasts (group advancement, title odds) come from Monte-Carlo simulation of the remaining schedule using those same match probabilities.

Bookmaker odds contain margin (the “overround”) — the implied probabilities of all outcomes sum to more than 100%. Before any comparison we de-vig: each outcome’s raw implied probability is divided by the overround so the market’s opinion sums to exactly 100%. The model’s edge on an outcome is then

edge = model probability − de-vigged market probability

Market prices are archived closing odds from published, verifiable sources — every settled pick links its odds source so you can check the price we graded against.

Three market families are priced per match: the 1X2 result, goal totals (Over/Under) at the market’s main line — among half-goal lines quoted by at least five bookmakers, the one whose consensus Over price sits closest to evens (never a fixed 2.5) — and both teams to score. Since 24 July 2026 the reference for these goal markets is a live multi-book consensus: the median price across every bookmaker quoting that exact line, with at least five books required before a pick can publish. Where no sharp bookmaker prices a market at all (BTTS in some leagues), the reference is a median of soft books only — a weaker benchmark, and the pick card says so (“soft-book reference”).

A pick exists if and only if, at the best-edge outcome:

• the edge is at least +2.5 percentage points against the de-vigged market, and
• the expected value at the raw odds is non-negative: EV = p × odds − 1 ≥ 0 (edge is measured after removing the vig, so a short-priced favourite can clear the edge gate while still losing money at the actual odds — those are never staked).

At most one pick per market per match (1X2, O/U main line, BTTS), and only corroborated odds may publish one — for goal markets that means a consensus of at least five books quoting the exact line. Confidence labels on the pick itself are mechanical, not editorial: HIGH when edge ≥ 6 points (or ≥ 4 points with model probability ≥ 45%), MEDIUM when edge ≥ 4 points, LOW otherwise.

• Two staking eras, boundary fixed in advance. Picks published before 15 July 2026 were staked a flat 1 unit. From 15 July 2026, every pick stakes the ½-Kelly percentage below applied to the record’s live bankroll at publish — 100 units at the start, plus all settled profit and loss, minus stakes still at risk in open picks — so winnings compound, as Kelly intends. Both the percentage and the resulting stake in units are frozen on the pick row at publish and reproducible from the ledger. No historical pick was ever re-graded.
• Stakes and odds are frozen when the pick is published — never revised at prices or probabilities seen later. No martingale, no discretion.
• Same-match picks share one exposure cap. A 1X2, an O/U and a BTTS pick on the same match are correlated — they often win or lose together — so the total stake at risk on any one fixture is capped at 10% of bankroll. Later picks in the publish order are scaled down, possibly to zero, and marked as capped on the card.
• Won → profit = stake × (odds − 1). Lost → −stake.
• Settlement is on the 90-minute score, including knockout games — extra time and penalties never count, matching how 1X2 markets settle everywhere. A knockout match that is level after 90 minutes settles as a draw.
• ROI = total units won or lost ÷ total units staked.

The stake itself — quoted on every open pick card and, since 15 July 2026, the record’s grading basis — is a deliberately humble Kelly criterion, with all three knobs fixed in advance: the model’s probability is first shrunk halfway toward the de-vigged market price, the stake is half of the Kelly fraction that shrunk probability implies, and it is capped at 10% of bankroll.

stake = ½ × (p′ × odds − 1) / (odds − 1), p′ = (model + market) / 2

Since 15 July 2026 both numbers are stamped on every pick row at publish, frozen alongside the rest: the percentage is re-derivable from the hash-committed probability and odds, and the stake in units from the ledger’s state at the publish timestamp — so the staking can be verified after the fact just like the pick itself.

Why not full Kelly on the model’s own probability? Picks are by construction the fixtures where the model most disagrees with the market — which selects its most overconfident probabilities. Sizing by those would over-stake ruinously even on a profitable pick set. The record page’s staking analysis shows the comparison, recomputed live from the settled ledger.

• Picks are published before kickoff, with the kickoff time on the card.
• Every pick's exact numbers (selection, model probability, odds, publish timestamp) are committed with a SHA-256 hash at publish time — anyone can recompute it.
• The ledger is append-only: settled picks are immutable and losses stay on the books forever.
• The published track record covers published picks only. Research backtests are never folded into the headline numbers.

To verify a pick yourself, follow the verification guide.

• The model can be wrong — sometimes badly, sometimes systematically. A published edge is an estimate, not a fact.
• International football is a small-sample sport; tournament records are short and variance is enormous. A good process can lose and a bad one can win over a World Cup.
• We hold one odds snapshot per pick (the archived closing price), so closing-line-value analysis is not yet part of the record.
• The goal markets have losses on the books and no proven edge.O/U and BTTS picks were retired on 13 July 2026 after a graded backtest lost 20.15 units over 115 bets, and relaunched on 24 July 2026 under a different regime: live multi-book consensus odds instead of researched files, calibrated per-market model probabilities, the same gates, and market-shrunk half-Kelly stakes. The prop-era losses remain in the ledger permanently. No market — 1X2 included — will carry a proven-edge claim until the model side’s closing-line-value gate for that market (at least 50 graded bets beating the close by ≥2%) reads MET; until then, the public ledger is the live test, not evidence of an edge.

None of this is betting advice — see the Terms and gamble responsibly, if at all.