When our model disagrees with the market

Last week we shipped a small panel on every match-detail page called Model vs Market. It puts our calibrated 1×2 probability next to the bookmaker's implied probability (after stripping the margin) and shows the difference in percentage points.

We hesitated for a while before building it, because this feature has an obvious failure mode: written badly, it becomes a value-bet finder, which is not what MatchMind is. Written carefully, it's the most transparent thing a football probability model can possibly do. Here's how to read it.

What the panel shows

Three rows: Home win, Draw, Away win. For each, the model's percentage, the market's implied percentage, two thin grey bars (same colour — the delta is the story, not which side is “winning”), and a signed Δ in percentage points. We also surface which bookmaker the odds came from and how much of the implied probability total was margin (the bookmaker's commission, removed before display).

That's it. No green-for-positive / red-for-negative colouring. No leaderboard of biggest divergences. No filter for “show me where the model has an edge.” The visual is deliberately boring so the visitor focuses on the numbers rather than the implied call to action.

What divergences actually mean

If our model says 52% home win and the market implies 48%, what should you conclude? Four honest possibilities, in roughly descending order of likelihood:

1. The bookmaker's margin is doing some of the work. Even after we normalise the implied probabilities back to sum to 100%, the relative weighting still carries some sportsbook-specific bias. The same fixture priced at two different books often gives slightly different implied probabilities. A 4-point gap is well inside the noise floor.
2. News asymmetry. A late team announcement (an injury, a tactical change, a manager press-conference quote) might be priced in by the book within minutes; our model only updates on Sportmonks pulls. If the gap appears after kickoff -60 minutes, suspect news.
3. Market efficiency varies. Big Premier League games have enormous liquidity — implied probabilities converge to the truth fast. A midweek Ligue 1 fixture between two mid-table sides has thinner liquidity and wider opportunity for the market to be wrong (or for us to be wrong).
4. Genuine model edge. Possible. Not provable from a single fixture. You need a long out-of-sample evaluation to distinguish edge from noise.

Notice what's missing from this list: any version of “back the model when it disagrees with the bookies.” That's a tipster move and it has eaten more amateur bankrolls than any other single mistake. Without out-of-sample calibration evidence, you can't know which of the four explanations above is driving the gap on any given fixture.

What the panel is for, then

The panel exists to show you that we're willing to be measured against the market. That's a high bar — most football probability sites bury their methodology specifically because they don't want their numbers compared. Putting our probabilities next to a bookmaker's is the strongest signal we can send that we mean it about calibration.

Use it as a thinking aid. If our model is 8 points off the market on a fixture you know well, that's a useful prompt to ask: what does our model see that the market is discounting? Or — much more likely — what does the market know that our model is missing? Either answer is interesting. Neither is a call to action.

What we explicitly will never build on top of this

• A filter for “fixtures where the model disagrees with the market by > X points”
• A sorted list of biggest divergences
• A “value bet alert” email
• Any feature that nudges the user from observing the gap to acting on it

Those features are easy to build. They're also exactly the path that turns a transparent calibration project into a tipster site. We're not doing that.

See it on a real fixture

Open any upcoming match in the fixtures listand the panel renders if we have odds for that match (we don't always — odds coverage is fixture-by-fixture). The methodology footer in the panel itself is the canonical text on what each number means.