One model, graded in public

"EV" is expected value: what a price should be worth against what it costs. Finding it is the easy claim to make and the hard one to prove, so this page explains what we actually build, what we feed it, and the two ways we let you catch us being wrong.

The model

A hand working through equations on paper in blue ink

For each match the model produces a single probability distribution over the score — not a prediction of who wins, but a shape describing how likely every margin and total is. Both markets we bet are read off that one object. That has a useful consequence: our margin and totals prices can never disagree with each other, because they're the same belief expressed twice.

Into it goes what you'd expect: years of results, team strength ratings, player-level statistics, venue characteristics, scoring-rate history, and the weather forecast as it stood at the time, never the version published after the game. Everything is fitted walk-forward: the model only ever sees seasons before the one it's being tested on.

What it can't do is worth stating plainly. Well under half the variance in a football margin is knowable before the bounce; the rest is the game itself. Any service that implies otherwise is describing a fantasy.

Scored properly

How it's scored

A model that outputs distributions can't be graded on "did it pick the winner". We use proper scoring rules: measures built so that the only way to score well over time is to report what you actually believe. Alongside them we check calibration. When the model says 60%, does the thing happen about 60% of the time?

A calibration curve (illustrative). Points on the diagonal mean the stated probabilities are honest. A model can have impressive-looking point estimates and still fail this chart badly — which is why we don't publish point estimates.

What lines do as the game approaches (illustrative): early prices wander, then converge on the close. We publish where our advised price sat against that close, tip by tip, in the ledger.

The market is the examiner

Bookmaker prices are a benchmark we score against, never data we train on. The closing line — the last price before play — is the strongest public estimate of a game that exists, because it has absorbed everyone's information including ours. So every tip is graded retrospectively against the close. Beat it consistently and the edge is real; fail to and no run of wins means much.

That grading is also the honest answer to "why totals and margins?" — they're two-way markets with published lines, which makes settlement mechanical and closing-line grading possible. Player props would be easier to sell and impossible to audit.

Discipline

Around ten tips in a normal week. When the model finds three prices worth taking, you get three. Volume targets corrupt records.
Level stakes, sized in units. Nothing in the ledger is staked to make a bad week look quiet.
Every constant in the pipeline has a documented origin, and any component that could read the future — a forecast revised after the game, a rating computed with hindsight — is found and removed before it ever prices a bet.
When we change the model, the record doesn't reset. The ledger is continuous or it's worthless.

Sports

The approach was built for AFL first. NRL and NBL run under the same settlement rules and the same grading; the A-League is under evaluation and joins only if the same discipline holds there. We'd rather cover three leagues honestly than six leagues thinly.

A football stadium under lights at night, seen from high in the stands

AFL

A bank of floodlights glowing through night mist

NRL

A basketball arena lit for a night game, the court empty

NBL

Roadmap — stated so you can hold us to it

A calibration page, published once the live sample is big enough to mean something.
Line-impact reporting: we track how far markets move after tips go out, and we intend to publish it — subscribers deserve to know what our own volume costs them.
Telegram delivery, after Discord has bedded in.