All the details

A retrospective replay of settled 1X2, over/under 2.5 and both-teams-to-score calls, using a notional $10,000 and flat 1% stakes. Prices must have been recorded at or before the prediction. That cannot be checked here yet: neither the table nor the export shows each price's time and bookmaker. This reconstruction is separate from the forward journal above; its balance is not carried into the new account.

Net P&L

+$1,385

Record

65W – 34L

Settled picks

99

Staking

Flat 1%

1% of bankroll per pick

Historical simulation, reconstructed from stored predictions and prices. It does not show bets actually placed or a real betting-account balance. It applies results in fixture order, without modelling overlapping open stakes or actual settlement times; it is not necessarily what a follower with limited available cash could have earned. The separate forward journal reserves stakes and records later settlements. This replay does not establish an edge. Past results do not guarantee future returns. 18+.

The historical AI bankroll replay above is one number, and one number can hide what made it: a book that only profits on odds-on favourites reads exactly like one that only profits on longshots. These are the same 99 settled bets, split by the price we took and by the market. Hypothetical, like every figure here.

A fixed 1 unit per bet across 99 settled bets, at the best screened price recorded at or before the forecast. Each price's time and bookmaker are not shown yet, so that cannot be checked here. A price band nobody bet in is left out rather than shown as a flat zero. 4 of 103 settled predictions had no price at the forecast and are left out.

Every signal whose match has a logged closing line is scored once it kicks off: the best screened price at detection vs the de-vigged closing consensus. Two caveats we would rather state than bury — it mixes line-shopping with timing, and it divides a price that still carries the bookmaker's margin by a margin-free fair price, so it reads low before any timing skill is involved. If the edge isn't there, this number says so.

Part of that number is the market's margin, not timing: on the 125 graded signals whose closing board stored a booksum, a bet struck at the market's own margin would have scored roughly -4.8% on its own. Net of that, the signals scored roughly +1.1% — which still includes whatever taking the best price across books earned, so read it as an upper bound on timing, and the replayed null below as the real comparison. Both figures are approximations: here the margin counts the same for every outcome, while the fair close puts more of it on longshots.

The null has graded nothing yet — last run 12 Oct 2026, with nothing to grade. It is replayed weekly.

• Arb · 10 settled · 60% beat close · -1.6% mean CLV · longest leg only — not a bet
• Model gap · 125 settled · 35% beat close · -3.7% mean CLV

See the live signal feed

When a match has a market price, the headline 1X2 probability is anchored to that vig-free price, and the model may lean at most ±3 pp per outcome. The price comes from the sharp bookmakers that quote the match — that can be a single one — or else from the median of all bookmakers. The rows are scored on the same live, market-anchored predictions: the model's own number before anchoring and the market price it was anchored to — both stored when the prediction was made — plus the closing line on those of the rows with a logged close. The live row is the anchored number we actually published. Every row carries its own n. We publish the table even where the market wins.

Brier = mean over the three outcomes, Σ(p−o)²/3; 0 is perfect, 0.667 is worst. Lower is better for both Brier and RPS.