Accuracy records how often a predicted outcome wins. Calibration asks whether events assigned a given probability occur at roughly that rate. Betting decisions need both probability quality and price context, so raw winner percentage alone is an incomplete model standard.
Accuracy answers only one question
Classification accuracy asks how often the predicted winner actually won. That can be informative, but it ignores the price required to make the selection. A model that always chooses a -300 favorite may post a strong win percentage while still producing little value—or losing money—if the winners do not occur often enough to justify the cost.
Sports betting is a pricing problem rather than a pure winner-picking contest. The model must estimate probabilities well enough to compare them with the market, and the available payoff must compensate for the risk.
A 70% winner is not automatically a good bet. The decision depends on whether the market price requires it to win more or less than 70% of the time.
Calibration compares forecasts with observed frequency
A model is well calibrated when outcomes assigned a 60% probability occur approximately 60% of the time over a sufficiently large and representative set of predictions. The same principle applies across probability ranges: 52% forecasts should not behave like 70% events, and 75% forecasts should not behave like coin flips.
Calibration is commonly reviewed by grouping predictions into probability buckets and comparing the average forecast with the observed result. The buckets must contain enough observations to be meaningful, and the review should account for sport, market, season, and other conditions that could hide different behavior inside one aggregate number.
| Forecast bucket | Average forecast | Observed win rate | Calibration gap |
|---|---|---|---|
| 50%–54.9% | 52.4% | 51.7% | -0.7 pts |
| 55%–59.9% | 57.2% | 56.5% | -0.7 pts |
| 60%–64.9% | 62.1% | 58.8% | -3.3 pts |
| 65%+ | 68.4% | 67.9% | -0.5 pts |
Calibration and discrimination solve different problems
Calibration measures whether the probabilities have the right scale. Discrimination measures whether the model meaningfully separates stronger outcomes from weaker ones. A model could be calibrated overall while assigning nearly every event 50%, which would provide little practical ranking power. Another model could rank outcomes correctly but be systematically overconfident.
Useful evaluation therefore considers both. The model should distinguish between events with different likelihoods and attach probabilities that behave realistically. Proper scoring rules such as log loss or Brier score can help evaluate the full probability distribution rather than reducing every forecast to a binary correct-or-incorrect result.
(forecast probability − actual outcome)², where the outcome is 1 for a win and 0 for a lossLower average Brier scores indicate better probability accuracy, but comparisons should use the same event set and context.Price determines whether a calibrated forecast has betting value
Even a well-calibrated model does not automatically create profitable wagers. If the sportsbook price already reflects the same or a stronger probability, there may be no positive expected value. Conversely, a small but well-supported difference can matter when the price is favorable.
That is why SLS evaluation separates model quality from recommendation performance. Calibration helps assess the probability estimates. Market comparison measures the disagreement with the price. Decision gates determine whether the resulting opportunity is publishable. Public performance then records what actually happened to official recommendations.
Sample design can change the conclusion
Calibration measured over a tiny sample can swing dramatically. Aggregating every sport and market can also hide systematic weaknesses, such as totals being overconfident while moneylines are well calibrated. A credible review should evaluate supported markets separately and combined, preserve the production decision path, and avoid selecting only the strongest-performing slice after seeing the results.
Time matters as well. League rules, scoring environments, rosters, market efficiency, and data quality can change. Monitoring should distinguish temporary variance from a persistent structural problem and should compare candidate changes on the same games and prices whenever possible.
- Report sample size and the dates represented.
- Evaluate each supported market as well as the combined output.
- Avoid retroactively redefining which historical predictions count.
- Compare model versions on matched events and prices.
- Keep published customer performance separate from research replays.
Limitations and interpretation
- The calibration table is hypothetical and is not an SLS historical-performance claim.
- Probability metrics can be unstable in small or non-representative samples.
- No single score captures calibration, ranking ability, price sensitivity, and realized betting performance at once.
Published by Sports Line Signal. This public research note is educational analysis, not a guarantee of outcome or a substitute for the live line, price, and recommendation context shown in the SLS product. 21+.