A betting model can report 100 predictions, a 60% win rate and an impressive historical return without having produced 100 independent pieces of evidence. When multiple predictions depend on the same games, players or market conditions, their outcomes may be correlated.
This creates a problem for anyone evaluating betting-model performance: the number of recorded bets is not necessarily the same as the effective sample size.
Effective sample size measures how much independent statistical information a dataset contains. It can be much smaller than the number of observations when outcomes move together.
That distinction matters because analysts sometimes use sample size to judge whether an apparent advantage is credible. A record with 100 correlated selections can provide considerably less evidence than 100 genuinely independent observations.
Why Multiple Betting Results May Tell the Same Story
Imagine a hypothetical analyst tracking five different predictions for each NFL game.
The predictions concern the moneyline, total points, quarterback passing yards, receiver yardage and another player-performance market.
Across 20 games, the analyst records 100 separate predictions.
Each prediction may have its own probability estimate, quoted odds and settlement result. However, several predictions from one game may depend on shared conditions.
An unexpected quarterback injury could affect the team’s chance of winning, passing production and receiver yardage simultaneously. A major change in weather could influence several scoring-related markets.
These results are not necessarily independent simply because they appear on separate betting tickets.
Statisticians call observations that share common influences clustered data.
The CDC’s guidance on variance estimation explains that observations within a cluster can resemble one another, reducing the effective sample size and making conventional uncertainty estimates misleading. Although the CDC discusses survey research, the underlying statistical principle also applies when evaluating groups of related betting outcomes.
This does not mean all selections from the same game are positively correlated. Some may move in opposite directions, while others have little meaningful relationship.
The correct treatment depends on the outcomes being evaluated and their actual dependence structure.
For a related discussion of dependence in pricing, BettorSearch’s guide to correlation in same-game props and parlays examines why combining related outcomes requires more than multiplying individual probabilities.
The additional problem for performance analysis is that correlation can also distort how confidently a historical record should be interpreted.
A 100-Bet Example: How Effective Sample Size Changes
A simplified calculation shows the potential effect.
Suppose an analyst has the following hypothetical dataset:
- 100 individual predictions.
- 20 different games.
- Five predictions associated with each game.
- An assumed average within-game correlation of 0.30.
For an idealized equal-sized cluster model, statisticians use a measure called the design effect.
The calculation is:
Design effect = 1 + (m − 1) × ρ
Here, m is the number of observations per cluster, and ρ represents their common within-cluster correlation.
Using five predictions per game and an assumed correlation of 0.30:
Design effect = 1 + (5 − 1) × 0.30 = 2.20
The effective sample size is then:
100 ÷ 2.20 ≈ 45.5
Under these assumptions, 100 recorded predictions provide approximately the same precision for an estimated average as 45.5 independent observations.
That is less than half the reported number of predictions.
The following table shows how changing only the assumed within-game correlation affects the estimate.
| Assumed within-game correlation | Design effect | Approximate effective sample size |
|---|---|---|
| 0.00 | 1.00 | 100.0 |
| 0.10 | 1.40 | 71.4 |
| 0.30 | 2.20 | 45.5 |
| 0.50 | 3.00 | 33.3 |
These are hypothetical calculations, not findings from actual betting records.
The mathematical approach is described in the PracTools documentation on design effects and effective sample size, which explains how within-group similarity changes the precision of statistical estimates.
The simplified formula assumes equally sized clusters, a common within-cluster correlation, comparable observation variances and independence between clusters.
Actual betting data rarely satisfies every assumption perfectly.
Different selections may have different probabilities, payouts, stakes and correlation structures. Games occurring on the same day can also share broader influences.
Consequently, the example should be understood as a demonstration of the problem, not a universal correction that can be applied to every betting record.
How Correlation Can Make a 60% Win Rate Look More Reliable Than It Is
Suppose the same hypothetical 100-prediction record contains 60 correct predictions.
Its observed win rate is 60%.
That number is arithmetically correct.
What changes is the uncertainty around the estimate.
Under a simplified independent-observation model, the estimated standard error of a 60% sample proportion is:
√(0.60 × 0.40 ÷ 100) ≈ 0.049
That is approximately 4.9 percentage points.
If a suitable clustered model instead produces an effective sample size near 45.5, the corresponding illustrative calculation becomes:
√(0.60 × 0.40 ÷ 45.5) ≈ 0.073
That is approximately 7.3 percentage points.
The observed win rate remains 60%, but the uncertainty is substantially greater.
These figures are illustrative standard-error calculations under the stated assumptions, not formal confidence intervals or measured results from a betting strategy.
The difference matters when someone claims a particular win rate demonstrates predictive skill.
A relatively small difference between an observed win rate and a relevant market benchmark may be difficult to distinguish from ordinary variation, especially when observations are dependent.
The situation becomes more complicated when different selections have different odds.
A 60% win rate cannot establish profitability without considering prices and settlement conditions. Winning 60% of very short-priced selections can still produce losses.
Likewise, a favorable historical return does not establish that the same outcome can be expected in the future.
BettorSearch’s guide to key betting performance metrics provides broader context on win rates, returns and odds. Correlation adds another consideration: how much independent evidence supports those measurements?
Why Standard Confidence Intervals Can Understate Uncertainty
A common statistical mistake is to calculate uncertainty as though each recorded prediction were unrelated to every other prediction.
When results are positively correlated within games, this approach can make the standard error too small.
That can produce confidence intervals that look more precise than the evidence justifies.
There are several ways an analyst can address this.
One is to calculate cluster-robust standard errors.
These methods adjust estimated uncertainty to account for dependence among observations within defined groups.
The Python statistics package statsmodels provides a cluster-robust covariance function for regression analysis. Its documentation also describes a small-sample correction and different clustered covariance estimators.
However, the existence of a statistical function does not automatically make an analysis valid.
The analyst must select meaningful clusters, use an estimator appropriate to the outcome and consider the number of independent groups.
Twenty games with five observations each should not automatically be treated as providing the same statistical information as 100 unrelated games.
For model-performance estimates, another possible approach is cluster-based resampling.
Instead of randomly resampling individual predictions, the analyst resamples entire games while keeping related predictions together.
This can better preserve within-game dependence.
It still requires appropriate assumptions. If predictions across different games are also dependent, resampling individual games may not capture all the uncertainty.
For example, several games played during the same week might be affected by shared market conditions or a common change in model inputs.
The right clustering level depends on what produced the dependence and which performance quantity is being estimated.
The Validation Problem: Related Predictions Can Appear in Both Datasets
Correlation also matters when separating training data from testing data.
Imagine a model trained to predict NFL player props.
An analyst randomly divides individual player-market records into training and testing groups.
Some predictions from a particular game enter the training set, while other predictions from that same game enter the test set.
This can create a misleading evaluation if information associated with the game appears in both groups.
The model might benefit from patterns or shared information that would not have been independently available when predicting an entirely new game.
A better validation design keeps observations from the same game together.
The scikit-learn GroupKFold documentation describes cross-validation that prevents the same group from appearing in both the training and test portions of a split.
For a sports model, a fixture identifier can be used to define these groups.
However, ordinary grouped cross-validation is not automatically chronological.
If the objective is to forecast future games, the analyst should also ensure training observations occur before testing observations.
The strongest approach may therefore require both chronological ordering and group separation.
This keeps related predictions together while preserving the direction of time.
What a Credible Betting Performance Report Should Include
When evaluating a model, prediction service or published betting record, the headline sample size is only a starting point.
A more informative report should disclose the following:
1. Number of predictions and distinct events. A record containing 1,000 selections across 100 games is structurally different from one containing 1,000 selections across 1,000 unrelated events.
2. Distribution of predictions across events. A few games containing unusually large numbers of predictions can make the overall record heavily dependent on those games.
3. Market relationships. Moneylines, totals, spreads and player props can share underlying drivers. The report should acknowledge relevant dependence without assuming every pair is correlated.
4. Probability and pricing differences. A simple win-rate analysis is insufficient when the selections have materially different odds, payouts or settlement conditions.
5. Uncertainty estimation method. The report should explain whether it assumes independent results, uses cluster-adjusted methods or applies another justified approach.
6. Independent testing. It should distinguish historical fitting from later validation, while explaining how observations from the same event were handled.
7. Complete financial records. Reported returns should use actual or properly qualified historical prices, consistent stake accounting, settlement outcomes and applicable costs.
These details are particularly important when a provider advertises statistical confidence or uses phrases such as proven accuracy and verified model performance.
A larger record may provide stronger evidence, but only if its observations and evaluation procedures support that interpretation.
More Bets Do Not Necessarily Mean More Evidence
Effective sample size changes how analysts should interpret betting records.
The hypothetical 100-prediction example demonstrates that moderately correlated selections can provide far less independent information than their raw count suggests.
That does not automatically make a model inaccurate or invalidate its recorded outcomes.
It means the statistical evidence may be weaker than a simple count implies.
A credible evaluation should preserve the number of predictions while also explaining their dependence, price differences and uncertainty.
The most useful question is not merely how many bets or predictions were recorded. It is how much new information each additional observation actually contributes.
No amount of statistical adjustment guarantees accurate predictions, profitable betting or reliable income. Proper analysis makes uncertainty more visible rather than eliminating it.