Definition
Regression to the mean arises when repeated measurements contain both a persistent component and temporary variation. Selecting observations because they are extreme also selects unusually large temporary components, so later measurements can be less extreme even without intervention. Barnett, van der Pols and Dobson explain this selection and measurement-error mechanism.
Football example
Suppose a team scores 12 goals from chances assigned 7.4 total xG across five matches. The finishing difference is:
12 - 7.4 = 4.6 goals
That does not prove future scoring will fall, because xG is model-dependent and the team, opponents and shot quality can change. It does warn that selecting the team for an extreme finishing gap can overstate the evidence for a stable new level.
Betting-record example
A method wins 18 of its first 25 even-money decisions. Observed strike rate = 18 / 25 = 0.72, or 72%. The result is extreme relative to 50%, but it does not establish a 72% future probability. The method must be evaluated through pre-event probabilities, calibration, selection rules and later untouched data.
Regression is not the gambler's fallacy
| Statement | Assessment |
|---|---|
| “An extreme noisy rate may be less extreme next period” | Possible regression-to-mean reasoning |
| “Five losses mean the next bet must win” | Gambler's fallacy |
| “The striker will decline because xG says so” | Unsupported causal claim |
| “We selected extreme finishers, so we need later confirmation” | Appropriate design caution |
Design a valid test
- Define the extreme selection rule using an earlier period.
- Preserve the comparison group, not only selected teams.
- Keep metric and provider definitions stable.
- Evaluate the next period without changing the threshold.
- Control changes in opposition, minutes, role and lineup.
- Report uncertainty and all selected cases.
For probability models, calibration guidance helps assess whether forecasts at a stated level resolve at a similar frequency. A short winning or losing run should not override the broader reliability record.
Practical use
Regression-to-mean awareness is a brake on dramatic stories. It encourages shrinkage, larger contextual samples, later confirmation and humility about extreme observations. It is not itself a betting signal.
Avoid the before-and-after trap
Selecting the hottest teams or highest-return tipsters and comparing their next period with the selected extreme creates an asymmetric setup. Build a comparison group using the same selection date, league, market, and exposure, then estimate how much persistence exists across all candidates rather than only the extremes. Regression-to-the-mean research explains why selection on an extreme measurement requires this control.
Where possible, use partial pooling so short records move toward a broader prior in proportion to their information. Report both raw and adjusted estimates and validate the adjustment on later data. A less extreme later result can be compatible with regression to the mean, a genuine change, or both; the repeated-measurement limits are described by Barnett and colleagues.
Next step
Use Regression To Mean for the next part of this topic.
Continue learning
- Next guide: Soft vs Sharp Bookmakers
- Related guide: Value Betting vs Matched Betting
- Definition: Regression to the Mean
Assumptions and limitations
The figures are illustrative. The page does not assert that goals must converge to xG or that every extreme record is luck. Persistent skill, tactical change, injuries, schedule and measurement changes can all alter the underlying level.
GambleAware
GamCare
Gambling Therapy
GAMSTOP