Regression to the mean is the tendency for a repeated noisy measurement to be less extreme after cases were selected for an unusually high or low first measurement. It is not a force that reverses results and it does not make a team or player "due" to improve or decline. Barnett, van der Pols, and Dobson explain that the effect is especially relevant when follow-up analysis selects a sample using an extreme baseline value and when measurement error is substantial.
The Basic Principle
As an illustrative sampling example, suppose many players have a stable underlying shot-conversion probability, but the observed rate over 20 shots contains sampling noise. If analysis selects only players who converted 8 of 20 shots, the group was selected at an extreme observed rate of 40%. On a later 20 shots under the same conditions, the group average is likely to be less extreme because the selection captured both underlying skill and favourable first-sample variation. This applies the repeated-measurement conditions described by Barnett, van der Pols and Dobson; it is not a claim about a documented player sample.
That statement depends on the conditions remaining comparable. A player with a new role, different shot locations, an injury, or a genuinely changed ability may have a different underlying process. The relevant comparison is therefore not automatically a career or league average.
How It Applies to Football Form
Form tables report recent outcomes; they do not isolate ability, opposition strength, venue, or sampling variation. The need to distinguish an extreme baseline selection from real change follows the framework in Barnett and colleagues' regression-to-the-mean paper. To assess a possible regression effect, define all four parts:
- the measurement, such as points per match, goals per shot, or xG difference;
- the rule that selected the extreme sample;
- the comparison mean for a genuinely comparable population or model; and
- evidence that the underlying process is sufficiently stable for a repeated measurement, because process change weakens a simple repeated-measurement comparison.
Expected goals can add information about chance quality, but it is another modelled measurement rather than a guarantee of future results. Opta's xG explanation shows that xG assigns a probability to shots from model inputs; provider definitions and model versions can differ.
Practical Football Example
Consider an unnamed, illustrative team selected because it won five consecutive matches. Its first-sample figures are:
| Metric | Selected five matches | Earlier comparison period |
|---|---|---|
| xG per match | 1.2 | 1.3 |
| Goals scored per match | 2.4 | 1.3 |
| Shot conversion | 22% | 11% |
The table shows an extreme realised scoring rate relative to two reference measurements. It does not prove that 11% or 1.3 is the current truth. Before forecasting, check shot locations and definitions, opponent strength, lineup and tactical changes, and uncertainty in both periods. A model may shrink the five-match estimate toward a longer-run estimate, but the amount of shrinkage must come from the model and data, not from the phrase "regression to the mean". Peer-reviewed football model comparisons show why feature choice and evaluation method remain empirical questions.
The Gambler's Fallacy Distinction
The gambler's fallacy says an independent outcome is due to compensate for previous outcomes, such as expecting heads because a fair coin produced several tails. Regression to the mean instead concerns repeated noisy measurements selected for being extreme, as set out by Barnett and colleagues. The next measurement can be closer to the mean without any compensating mechanism, and it can also be equally or more extreme.
Implications for Betting
Use the concept as an audit question, not a selection rule:
- Was the case chosen because its first measurement was extreme?
- How noisy is that measurement at the observed sample size?
- What baseline is relevant, and how uncertain is it?
- Has the process or context changed between measurements?
- Does an out-of-sample model improve on simply using the longer-run average?
- Does the available price compensate for the forecast uncertainty and market margin? A probability forecast still requires separate model evaluation, as demonstrated in football outcome-model comparisons.
Past performance does not guarantee future results. Regression to the mean is a probabilistic tendency, not a certainty for any individual match.
18+. Gambling involves risk. Please gamble responsibly. If you need support, visit begambleaware.org.
Related resources
Continue with Expected Goals (xG): What It Means in Betting for the next part of this topic, or return to Betting Glossary: Every Betting Term Explained in Plain English to compare the other guides in this collection.
Continue learning
- Next guide: Scorecast
- Related guide: Stake
- Go deeper: Regression to the Mean in Betting Analysis
Assumptions and limitations
Regression to the mean requires repeated noisy measurements and becomes especially relevant when an extreme observation was used to select the case. It does not mean every streak must reverse or identify the long-run mean by itself. The cited Barnett, van der Pols and Dobson paper supports these conditions; football examples remain illustrative.
GambleAware
GamCare
Gambling Therapy
GAMSTOP