Closing-line value (CLV) measures whether a football prediction model found an edge against the market’s final, most informed odds. If a model consistently assigns higher probabilities to outcomes than the closing line implies, it is said to have positive CLV—a sign of predictive skill that complements metrics like Brier score, log loss, calibration, ROI, and maximum drawdown. In AlphaBench, CLV is used as one benchmark among several, not as betting advice.
Why CLV matters for benchmarking
The closing line is the last set of odds available before kickoff, after the market has absorbed team news, weather, travel, and public money. Because it reflects a broad consensus, it is often treated as the market’s best estimate of each outcome’s probability. A model that beats the closing line is finding information or structure the market missed.
For benchmarking, CLV is valuable because it tests a model against a strong, real-world baseline. Accuracy alone can be misleading: a model might correctly predict many favorites, but if the closing line already priced those favorites heavily, the model adds little. CLV asks a sharper question: did the model’s probabilities differ from the closing line in a way that later proved useful?
AlphaBench tracks CLV alongside other metrics such as Brier score, log loss, calibration, ROI, and maximum drawdown. Each captures a different aspect of model quality. CLV focuses on market-relative edge, while the others measure accuracy, uncertainty, and risk. You can see how these fit together in our methodology and in how AlphaBench scores models.
How we measure CLV in the pilot
In the AlphaBench pilot, we record each model’s predicted probabilities for home, draw, and away at a fixed time before kickoff. We then capture closing odds from a reference market and convert them into implied probabilities, removing the bookmaker margin where needed.
CLV for a single outcome can be expressed two ways:
- Probability difference: model implied probability minus closing implied probability. Positive means the model rated the outcome more likely than the closing line.
- Odds ratio: (model fair odds ÷ closing odds) − 1, where model fair odds = 1 ÷ model probability. Positive means the model’s fair price was higher than the closing price.
We compute this for every fixture and aggregate across the pilot for each model, including examples like deepseek-flash (DeepSeek), gpt-5-mini (OpenAI), and grok-build-0.1 (xAI). In a recent settled fixture between Khovd Western and Deren (Premier League), we would calculate CLV for each outcome and then average or sum those values depending on the leaderboard view. The pilot is ongoing, so we do not report specific CLV numbers here; the goal is to show the method, not to make claims about any single match.
To illustrate without using real results: suppose deepseek-flash gave Khovd Western a 45% chance while the closing line implied 40%. The probability-difference CLV for that outcome would be +5 percentage points. If the model’s fair odds were 2.22 and the closing odds were 2.50, the odds-ratio CLV would be negative—showing that the two formulas can point in different directions when margin and probability scales differ. This is why AlphaBench reports the method clearly and uses consistent conventions. See methodology for details.
Limitations
CLV is a useful benchmark, but it has important limits:
- Closing lines are not perfectly efficient. Smaller leagues or lower-liquidity markets—like the one featuring Khovd Western vs Deren—may have noisier closing prices. A positive CLV there may reflect market disagreement rather than model skill.
- CLV is a proxy, not a guarantee. Positive CLV can occur by chance, especially with small samples. It does not promise future performance or profit.
- It depends on the reference market. Different books or odds providers can produce different closing lines. AlphaBench documents its source and timing to keep comparisons fair.
- It should be combined with other metrics. A model can show positive CLV yet have poor calibration or a high Brier score. Conversely, a well-calibrated model may show little CLV if the market is already efficient. Log loss, ROI, and maximum drawdown help complete the picture.
- Not betting advice. AlphaBench is an educational benchmarking project. CLV is used to evaluate models, not to recommend wagers.
Related reading
- AlphaBench methodology — how the pilot collects predictions, odds, and closing lines. - Brier score for football predictions — a core accuracy metric that pairs well with CLV. - How AlphaBench scores models — the full scoring framework, including calibration, log loss, ROI, and maximum drawdown. - AlphaBench leaderboard — compare model performance across the pilot, including CLV and other metrics.