The key idea: a bookmaker’s odds do not represent pure probabilities. They include a margin, often called the overround, which makes the implied probabilities sum to more than 100%. For football prediction benchmarks, the margin is the difference between raw market odds and the fair line you should compare model probabilities against. If you ignore it, you can overstate model error, misread calibration, and misjudge closing line value.
Calculating overround (1X2 example)
Decimal odds convert to implied probability with a simple formula:
\[ \text{Implied probability} = \frac{1}{\text{decimal odds}} \]
For a 1X2 market, add the three implied probabilities. The total will usually be above 1.00. The excess is the bookmaker’s overround or margin.
\[ \text{Overround} = \left(\frac{1}{O_H} + \frac{1}{O_D} + \frac{1}{O_A}\right) - 1 \]
Suppose, illustratively, a bookmaker prices the Premier League fixture Khovd Western vs Deren at:
- Home: 2.00
- Draw: 3.50
- Away: 4.00
The raw implied probabilities are:
- Home: \(1 / 2.00 = 50.00\%\)
- Draw: \(1 / 3.50 = 28.57\%\)
- Away: \(1 / 4.00 = 25.00\%\)
Total: \(50.00\% + 28.57\% + 25.00\% = 103.57\%\)
So the overround is 3.57%. To get margin-free, or fair, probabilities, divide each raw implied probability by the total:
- Home: \(50.00 / 103.57 = 48.28\%\)
- Draw: \(28.57 / 103.57 = 27.59\%\)
- Away: \(25.00 / 103.57 = 24.14\%\)
Fair odds are the inverse of those probabilities:
- Home: \(1 / 0.4828 = 2.07\)
- Draw: \(1 / 0.2759 = 3.62\)
- Away: \(1 / 0.2414 = 4.14\)
That is the basic margin-removal step. It does not make the market “true”; it makes the market probabilities internally consistent and comparable to model output.
Why margin matters for models
A football model usually outputs a probability distribution: for example, 48% home, 28% draw, 24% away. A bookmaker’s raw implied probabilities, however, are inflated by the margin. If you compare model probabilities directly with raw implied probabilities, you are comparing against a set that sums to 103%, 105%, or even more. That distorts evaluation.
Margin removal matters for several AlphaBench-style metrics:
- Brier score: measures the mean squared error of probability forecasts. If the reference probabilities are not margin-free, the benchmark is not a proper probability target.
- Log loss: penalizes confident wrong probabilities. Raw overround can make a model look worse or better depending on how the margin is distributed across outcomes.
- Calibration: checks whether events predicted at 40% happen about 40% of the time. Calibration against raw implied probabilities is misleading because those probabilities do not sum to 1.
- ROI: in a simulated evaluation, a model’s edge must be measured after the bookmaker’s margin. A 2% probability edge can disappear inside a 5% overround.
- Maximum drawdown: even a model with a real edge can experience long losing runs. Margin affects the size of that edge, so it affects drawdown profiles in any backtest.
In AlphaBench, models such as deepseek-flash from DeepSeek, gpt-5-mini from OpenAI, and grok-build-0.1 from xAI are evaluated as probability forecasters, not as betting systems. That distinction matters. The question is not “did the model beat the price?” but “how well did its probabilities match the margin-free outcome distribution?” You can read more about the scoring approach in How AlphaBench Scores Models and Brier Score in Football Predictions.
Margin varies by market and timing
The overround is not a fixed number. It changes by market, league, liquidity, and time.
A 1X2 market often carries a higher margin than a major over/under or Asian handicap market because it has three outcomes and more ways for the bookmaker to manage risk. Lower-liquidity leagues tend to have wider margins than heavily traded top-division fixtures. Outright and futures markets can have higher margins still because the bookmaker carries the position for months.
Timing matters too. Opening odds are usually set with less information and lower liquidity, so they often have a wider margin. As kickoff approaches, more money and information arrive, and the margin typically tightens. The closing line is usually the sharpest version of the market, but it still contains a margin.
That means a model evaluated against opening odds and a model evaluated against closing odds are not facing the same benchmark. Even for the same fixture — say Khovd Western vs Deren — the fair probabilities can move as team news, lineups, and market pressure arrive. AlphaBench-style evaluation should therefore compare like with like: same market, same timestamp, same margin-removal method. The methodology page explains why consistent data handling matters more than any single metric.
Margin and closing line value
Closing line value, or CLV, is a common way to measure whether a forecast contained information that the market later recognized. In simple terms, positive CLV means the price or probability you measured was better than the closing price or probability.
Margin removal is essential for CLV. Suppose a model gives Khovd Western a 50% fair probability. The closing home odds are 1.90. Raw implied probability is:
\[ 1 / 1.90 = 52.63\% \]
If the closing market has a 4% overround, the margin-free home probability might be closer to 50.6%, depending on how the margin is allocated. The model’s apparent gap shrinks. If you skip margin removal, you may think the model is far behind the closing line when the real difference is small — or you may overstate an edge that is mostly margin.
CLV is useful because it tests whether a model is learning something the market values. But it is not the same as calibration. A model can be well calibrated and still show no CLV if the market already knows what the model knows. A model can also show short-run CLV by luck. That is why AlphaBench-style scoring uses multiple views: probability accuracy, calibration, CLV, ROI, and maximum drawdown.
Limitations
Margin removal is an estimate, not a perfect recovery of truth. The most common method — dividing by the overround — assumes the bookmaker spreads margin proportionally across outcomes. In practice, bookmakers may load more margin onto longshots, favourites, or outcomes they fear. More advanced methods such as Shin, power, or odds-ratio models can produce different fair probabilities.
The closing line is also not an infallible oracle. It is a market consensus shaped by liquidity, limits, and information. It can be wrong. And CLV samples are often small, so a model that looks sharp over 50 fixtures may look ordinary over 500.
Finally, margin is only one input into evaluation. A model can be well calibrated on margin-free probabilities but still have poor ROI after real-world frictions. It can have positive CLV but a painful maximum drawdown. And none of this is betting advice — it is a framework for measuring forecast quality. For a fuller picture of how AlphaBench handles these trade-offs, see methodology.
Related reading
- AlphaBench methodology
- Brier score in football predictions
- How AlphaBench scores models
- AlphaBench leaderboard
Understanding overround turns raw bookmaker odds into a cleaner benchmark. It does not remove uncertainty, but it stops the margin from doing extra damage to model evaluation. That is the habit worth keeping: convert odds to margin-free probabilities before comparing them with any football prediction model.