MODEL EXPLAINER Β· 7 min read

How Bookmakers Set Odds (and Where the Margin Goes)

When you see odds of 2.50 on a home win, it is tempting to read them as the bookmaker's forecast: "the book thinks the home team has a 40% chance." That reading is only half right. Odds are prices, not predictions β€” and they carry a margin you never see, plus a market that keeps editing them until kickoff.

Here is the whole process in three steps: the opening number, the hidden margin, and the market's rewrite.

Step 1: the opening number

Someone has to post the first price. That job belongs to odds compilers at the biggest books, and their starting point is a statistical model much like GoalMath's predictor: power ratings for each team, adjusted for home advantage, injuries, rest, travel and recent form. Monte-Carlo simulations β€” thousands of virtual replays of the match β€” turn those inputs into fair probabilities for every outcome.

Because an opening line is just one firm's opinion, limits are kept low at first: the book is price-finding without risking much. If the early number is wrong, the first informed bets will say so loudly, and the line will move. This is the price discovery phase, and it is where the sharpest price changes of the week happen.

Step 2: the overround β€” the margin you never see

Bookmakers never publish fair odds. Every price carries a built-in margin called the overround (also "vig" or "juice"), and it is the engine of their business.

Take a fair market: home win 50%, draw 25%, away win 25%. Fair decimal odds would be 2.00 / 4.00 / 4.00, and the implied probabilities β€” just 1 Γ· odds for each β€” add up to exactly 100%. Nobody makes a penny on a fair market.

Now look at a typical published line: 2.10 / 3.40 / 3.60.

1 Γ· 2.10 = 47.6%
1 Γ· 3.40 = 29.4%
1 Γ· 3.60 = 27.8%
Total: 104.8% β€” the extra 4.8% is the overround.

That extra ~5% is the bookmaker's expected cut on every unit staked, whichever outcome wins. And it varies deliberately by market. Typical indicative ranges: top-flight match results around 4–6%, lower leagues 8–12%, over/under 2.5 goals 5–7%, correct score 15–25%, and exotic "bet builder" style markets 30% or more. Where casual punters dominate and prices are hard to compare, the margin swells; where informed bettors shop carefully, competition squeezes it.

The key insight: published odds systematically understate true chances. An implied 47.6% on a 2.10 price is not a fair 47.6% β€” it is a fair ~45.4% with the margin on top. Compare any model against market prices only after accounting for this.

Step 3: the market rewrites the price

From the opening line to kickoff, the odds live. Every bet nudges them β€” but not all money is equal. A €20 casual bet moves the line far less than a €20,000 bet from a bettor the book respects, because the book reads the big bet as information.

When money piles up on one side, the book shortens that price and lengthens the others β€” partly to rebalance exposure, partly because it trusts the informed money's judgment. A sudden coordinated move across several books is called a steam move; a line shifting against the majority of public bets often signals respected money on the other side.

The final price at kickoff β€” the closing line β€” has absorbed all the news and all the money. It is the most information-rich price of the week, and consistently the hardest single number to beat in all of sports modeling.

Closing line value: the one honest benchmark

This gives the industry its standard yardstick: closing line value (CLV). Compare the price you took with the closing line. If yours was better, you got positive CLV β€” you bought before the market finished its work, which is historically the mark of genuinely good pricing. If yours was worse, the market moved away from you.

CLV does not say whether a single bet wins. It says whether, across many bets, your estimates are better than the book's final number. No other simple metric separates signal from luck as well.

What this means for using our predictor

GoalMath's predictor outputs fair probabilities β€” they sum to exactly 100%. Market odds include the margin, so comparing them directly is apples to oranges. To compare fairly, strip the overround first: divide each implied probability by the total.

Fair: 47.6 Γ· 104.8 = 45.4% Β· 29.4 Γ· 104.8 = 28.1% Β· 27.8 Γ· 104.8 = 26.5%

That normalization is what our odds tools do behind the scenes, and it is the only honest way to ask "does my model disagree with the market?" β€” because the model estimates what the market estimates, before the book takes its cut.

One last sobering fact: the closing line usually beats individual models, including ours. The market aggregates thousands of opinions, models and news signals. A model like Dixon-Coles is a rigorous starting point for thinking β€” the market is a hard opponent to beat, and it is getting harder every year. That is worth knowing, and exactly why we publish our calibration record, including where we fail.

Related reading: How to Read Our Match Predictions Β· How the Dixon-Coles Model Predicts Soccer Scores