What 'Value' Means in a Football Prediction, and How It's Measured

Decimal odds of 2.00 imply a 50 percent chance; odds of 4.00 imply 25 percent. Add up the implied probabilities of every outcome in a bookmaker's 1X2 market and the total comes to more than 100 percent, typically 103 to 108. The excess is the margin, sometimes called the vig or overround, and it is the bookmaker's built-in profit.

Before a model's probability can be compared with a price, that margin has to be removed. Dividing each implied probability by the total, or using a method that accounts for favourite-longshot bias, gives de-vigged probabilities that sum to 100. Those are the market's real opinion, and they are usually the best public forecast of a football match that exists.

A selection has value when the model's probability is higher than the de-vigged market probability by enough to cover the margin and the model's own error. If the model says 45 percent and the fair market probability is 40, the gap is five points. The expected return on that selection at the actual price is the model's probability times the odds, minus one.

LiveWin's value finder flags a selection only when the de-vigged gap is at least 3.5 percentage points and the expected return at the captured price is above 2.5 percent. Below those thresholds a difference is treated as noise, because a model that is off by two points on a typical match is not unusual and a two-point 'edge' is more likely to be model error than market error.

The word is used loosely. A tipster who says a price is 'too big' is expressing an opinion with no probability attached, so the claim cannot be checked. A site that shows value badges without publishing its probabilities, the price it compared against, and the graded outcome is asking for trust, not offering evidence.

The other trap is timing. A price captured after team news or after money has moved is not the price a reader could have taken, and value measured against it is fiction. LiveWin freezes the best available pre-match price at the moment the forecast is recorded, and the return on the ledger is computed at that price, never at one found later.

The honest summary is that the model is reasonably calibrated and has not beaten the bookmaker margin. On the priced subset of the public ledger, a level-stake approach shows a negative return, and that figure is published on the accuracy page next to the hit rate. That is what a real value test looks like: a number that can go either way, updated daily, recomputable from the CSV.

It also shows why calibration and value are different achievements. A model can say 60 percent and be right 60 percent of the time while still failing to find prices where the market says 55. Value requires being both calibrated and different from the market in the right direction, and the market is a very hard opponent.