Model

Regression to the mean in football: why hot starts fade

By FootInsights · Published 16 Oct 2026 · 7 min read

A side opens a campaign scoring at nearly twice the league rate, and every preview written about it treats that rate as its identity. A few months later the rate has quietly collapsed towards ordinary, and the same previews explain it with fatigue, complacency or a tactical puzzle solved. Regression to the mean in football is the far duller explanation, and usually the correct one: the early rate was never the team’s ability, it was ability plus a run of luck, and only the ability part carries forward.

This is not a vague caution about small samples. It is a specific piece of arithmetic that tells you how much of an extreme number to believe, and it sits inside every serious prediction model whether or not the model’s authors name it.

What regression to the mean actually is — and what it is not

Start with the decomposition every rating model rests on. What you observe over a stretch of matches is not a team’s true rate. It is:

observed = true ability + sampling noise

Noise has no memory and averages to zero over the long run, but in any short window it is large. So the teams sitting at the top of an early scoring chart are disproportionately the teams whose noise happened to be positive. Their ability is above average — that is why they got there — but not by as much as the chart says. Going forward, the noise resets while the ability persists, and the observed number drifts down towards the mean.

The critical distinction: this is not the law of averages, and not the gambler’s fallacy. Nothing owes anyone a correction. A team that overperformed does not have to underperform to “balance out”. It simply performs at its true rate from here, and the cumulative average falls because the ordinary new matches dilute the extraordinary old ones.

The numbers make the distinction concrete. Suppose a team’s true rate is 1.4 goals per match, and across its first eight matches it scores 19 — a rate of 2.38. If it then plays the remaining thirty matches at exactly its true 1.4, it scores 42 more. Season total: 61 goals in 38 matches, or 1.61 per match. Well below the early 2.38, still above the 1.4 baseline, and at no point did it need a cold streak. That drift is regression to the mean.

The shrinkage formula, with a worked example

The practical question is not whether to discount an extreme number but by how much. The standard answer is a weighted average between what you observed and what you would have assumed knowing nothing:

estimate = (n × observed + k × league mean) ÷ (n + k)

n is the number of matches observed. k is the regression constant — the amount of league-average evidence the prior is worth, expressed in matches. When n = k, you split the difference evenly between the team’s own record and the baseline.

Take that same team: 2.4 goals per match over 8 matches, in a league averaging 1.4, with an illustrative k of 20 matches.

(8 × 2.4 + 20 × 1.4) ÷ (8 + 20) = (19.2 + 28.0) ÷ 28 = 1.69

The honest estimate is 1.69, not 2.4. Roughly seventy per cent of the apparent gap above league average was discarded as noise. Now run the same team forward to 30 matches at the same observed 2.4:

(30 × 2.4 + 20 × 1.4) ÷ (30 + 20) = (72.0 + 28.0) ÷ 50 = 2.00

Same observed rate, very different conclusion, because thirty matches of it is much harder to produce by luck than eight. Shrinkage is not scepticism applied uniformly — it is scepticism that decays as evidence accumulates, which is exactly the behaviour you want.

One more calculation shows why the discount at n = 8 has to be so aggressive. Goal counts are roughly Poisson, so the standard deviation of a team’s observed per-match rate over eight matches, at a true rate of 1.4, is √(1.4 ÷ 8) ≈ 0.42 goals. Line up twenty teams that are all identical at 1.4, and the highest of twenty draws lands around 1.9 standard deviations up — about 2.2 goals per match. A league of perfectly equal teams still produces a “runaway attack” averaging 0.8 goals above everyone else after eight matches. Any method that reads early tables at face value is guaranteed to find a story that is not there.

Not every football statistic regresses at the same speed

The regression constant k is not one number for all metrics, and this is the part that separates a useful model from a naive one. k is small — little shrinkage needed — when a statistic accumulates many events per match. It is large when the statistic depends on a handful of rare events.

The ordering matters more than any exact constant, and it explains why expected goals earned its place in modelling. An xG figure is built from every shot a team takes, so a single match contributes ten or more data points instead of one or two goals. More events per observation means less noise per observation, which means a smaller k and a faster-settling estimate. This is also why a team’s xG total and its actual goal total diverging is such a common signal: the noisier of the two numbers is the one that will move, and it usually moves towards the quieter one.

The failure mode to avoid is the reverse inference. A goalkeeper with an extraordinary save percentage over a short run is the single least reliable number on a stat sheet, and treating it as a durable property of the team is how models end up confidently wrong.

How prediction models apply regression to the mean

A well-built model does not bolt regression on at the end. It gets it structurally, in three places.

Pooling towards a league prior. Attack and defence ratings are estimated with the league’s own distribution acting as a prior. A team with a thin record is pulled hard towards average; a team with a full record barely moves. That is the shrinkage formula, applied automatically, with k implied by the fitted variance rather than picked by hand.

Recency weighting rather than windows. Old matches fade smoothly instead of being cut off, so an estimate is built from a large effective sample concentrated on the recent end — which is a far better trade between staleness and noise than reading the last five results. The mechanics of that curve are worked through in recent form in football predictions.

Opponent adjustment. Some of what looks like a hot start is simply a benign run of fixtures. A model that scores every result against the strength of the opponent removes that portion before any regression question arises.

The payoff is calibration. An unshrunk model reads 2.4 goals per match as 2.4, produces over-confident probabilities, and pays for it in its Brier score — the metric that punishes confident errors far harder than cautious ones. Shrinkage moves probabilities towards the middle, which feels like weaker analysis and is in fact the reason a calibrated model beats a bold one over a season. Whether a published model actually achieves that is not a claim to be taken on trust; it is what a running, dated record is for, and ours is on the track record page.

Reading regression correctly

Three rules follow, and none of them require running a model yourself.

1. Ask how many events, not how many matches. Eight matches of shot volume is real information. Eight matches of conversion rate is almost none. The number of underlying events sets the credibility.
2. Expect the drift without predicting the crash. The correct forecast for an overperforming team is ordinary from here, not bad from here. Backing a collapse is the gambler’s fallacy in disguise.
3. Apply it to records, too. A run of correct predictions regresses exactly like a run of goals. The only defence is a sample long enough that ability outweighs noise, which is the arithmetic set out in sample size in betting records.

Regression to the mean is unglamorous and it makes forecasts more boring. It is also the single cheapest correction available: no new data, no new features, just the discipline to treat an extreme number as partly luck in proportion to how easily luck could have produced it.