How it works

What the model does

Each sport has its own model. None of them “knows” the result of the game they are pricing. Ratings and features are built from information that was already available before that game started.

NBA

The model rates every player from recent box scores, weighted toward recent games. Before tip-off it reads the official injury report, works out the minutes each side is likely to play, and turns that into a team strength for tonight. That is blended with a long-run team rating, recent team strength, rest, back-to-backs, travel and a home-court edge that has declined over the years, to give each side a win probability. Super — best value, described below — is an NBA-only flag on top of that number.

NFL

The model rates every team from results and winning margins, adds a measure of how efficiently they move the ball (expected points), then home field and rest, and checks who is expected to start at quarterback. A backup or an inexperienced starter counts against that team. Picks can go up as early as a week before kick-off and update as quarterback news comes in, then freeze at kick-off.

NHL

Team ratings come from official results and shots on goal. They move after each game, shrink a little toward average between seasons, and a back-to-back is a small penalty. Every game is priced from earlier dates’ results only. The pick is which team wins the game, including overtime and shootouts.

Premier League

The model keeps attack and defence ratings from results and shots on target, and starts newly promoted sides a little below average. It publishes three chances — home, draw and away — and the pick is the most likely of those three. The settings were chosen on 2021–22 and 2022–23; later seasons are a clean test, as the backtest page sets out.

Championship

The same settings as the Premier League model, never retuned on Championship results. 2019–20 is used only to warm up the ratings. Again the pick is the most likely of home, draw and away.

Tennis

Every player has a rating that moves after each match, overall and on each surface (hard, clay, grass), built from more than 100,000 ATP and WTA results since 2005. Beating a strong player moves it more than beating a weak one; a retirement counts for half. The win chance blends the two players’ overall and surface ratings, their ranking, and whether the match is best of five. Picks refresh until the first serve. The final step of the model was fitted on 2009–2020 and was not shown the 2021–2026 matches used in the backtest.

How the lights work

Every pick gets a coloured light so you can see, at a glance, how strongly the model rates its own call. The cut-offs are not the same in every sport, and that is deliberate. We do not move them to chase a short run of results. The short page — How the lights work — is the reference. In brief:

Super is different. An NBA pick only gets Super when the model is at least 60% on the side, that chance is at least 3 points above the sharp market’s no-margin price, and the best Australian bookmaker price is at least 3% better than our fair price (with a cap on how large that gap can be).

If key players are listed as uncertain, a value pick is held back. It keeps its normal confidence light — so a 63% pick stays orange — and we add a note that it would have been Super except for those injuries. We do not promote a held pick to Super, because the Super test knew the real line-ups.

When picks are published

Every pick is timestamped before the start and can still update until then, as injuries, quarterbacks or prices change. It is frozen when the game begins (at first serve in tennis). The frozen chance is the one that goes on the record. Lights are not changed after the result.

The price we show next to a finished pick is the best Australian bookmaker price at our last check before the start — by 7am Sydney time on the day, and at least an hour before. “Vs close” compares that price with the fair closing price (the sharp market’s last price, with the margin removed). Above zero means the price you could get beat the market’s final view.

How results are graded

The Results page is the live record: only picks that were timestamped and frozen before the start. Green leads, but orange and red stay on the page. Nothing is deleted because it lost. You can download the CSV and check every pick yourself, including the posted time and the vs-close column where we have a price.

We also publish calibration — if the model says 70%, those picks should win about 70% of the time. Past-season tests of the same idea live on the backtest page. They are not added into the live green totals or the figures on the home page.

What we don’t claim

• These figures are for entertainment, not betting advice, and they are not a guarantee. A 70% pick is still meant to miss about three times in ten.
• A light is not a staking plan. Green is about confidence. Super is about price. Backing every green pick is not the same as backing Super.
• A backtest is a re-run on the past, not a record of tips we published at the time. It can look tidier than live results. That is why it has its own page.
• We do not claim to be the best, the most accurate, or to have a long public live record. The live site is recent. The backtests are how we tested the models before that.

Live results → · Backtest — not live results → · How the lights work → · FAQ →