Methodology transparency

Most betting products ask you to trust a number. This one tries to give you enough to check it.

Why this page exists

Content here is published under an organisation byline rather than an individual's name. That is a real trade-off: a named author with verifiable expertise is a stronger credibility signal, and choosing not to use one means giving up something.

The only honest way to compensate is to publish more of the method than a personal byline would require. What follows is that attempt.

What is published

How a grade is made. The full four-step sequence — sample, model, de-vig, publish the gap — with worked numbers, on how the edge is computed.

Which model prices which market. All 62 gradeable market keys with the distribution used for each, on markets covered. That table is generated from the engine's own registry rather than written by hand, and a parity check prevents it drifting from the code.

What data sits behind each league. Which sports have a dedicated warehouse with player logs and which run on universal scoring history, on sports and leagues covered — including the leagues where coverage is thinner than we would like.

What confidence measures. The three inputs that move it, the threshold that suppresses a grade entirely, and how it feeds back into shrinking the model toward the market, on the confidence score.

Where the model is weak. Thin early-season samples, the de-vig bias on lopsided markets, no weather or game-script variable, no adjustment for coaching behaviour in decided games, and no view of correlation across your own bets. All on known limitations.

What is missing from the product. The roadmap names the gaps between this and an actual betting assistant, including the ones that have not been started.

The commitments behind it

No fabricated track record. No historical performance claim appears anywhere until the calibration ledger has enough settled bets to be meaningful. This is the most commercially costly commitment on the list, and it is the one most worth keeping.

Refuse to grade rather than guess. Where the sample is too thin, the bet is marked THIN DATA. Where no model covers the market, UNRATED. Both are published rather than hidden, which makes coverage look worse and the product more useful.

Publish NEUTRAL. Most bets are fairly priced, and saying so is more valuable than manufacturing an opinion. A tool that finds a strong play in every game is not analysing, it is generating.

How to check any of it

The tests that do not require trusting us:

• Take a market, de-vig it yourself with the no-vig calculator, and compare to the fair price shown on the bet.
• Check whether the coverage tables match what the board actually grades.
• Watch whether THIN DATA and UNRATED counts rise in the first weeks of a season, as the stated method says they should.

If any of those do not hold, something on these pages is wrong, and corrections explains how that gets fixed.