Recent form in football predictions: how much should the last five matches count?
By FootInsights · Published · Updated 25 Sep 2026 · 6 min read
Every fixture list ships with a little row of coloured letters — W W D L W — and every preview builds an argument on top of it. Recent form in football predictions is treated as the decisive input: the team on a run is backed, the team on a slide is opposed. Yet ask how much the last five matches should count, in numbers rather than adjectives, and the confident answer usually evaporates.
The question has a real answer, and it is not “five matches”. It is a weighting curve, and the curve is one of the few genuinely tunable choices inside a prediction model.
Form is not a variable, it is a weight
The first misconception is structural. In most published models, “form” is not an input sitting alongside home advantage and attack strength. There is no column for it. A rating model estimates each team’s attacking and defensive rates from the results it has seen, and recency enters by deciding how much each of those results counts.
That is a subtle but consequential difference. If form were a separate variable, it could be added to a rating and the two would double count — the last five results are already inside the ratings that the last five results helped fit. Treating recency as a weight instead of a term avoids that trap entirely. A team in genuinely improving shape does not get a bonus on top of its rating; its rating simply rises faster than an unweighted fit would allow.
The weighting curve, in one number
The standard mechanism, proposed by Dixon and Coles in 1997 and still the reference implementation, is exponential decay: every historical match gets a weight that shrinks smoothly with its age. The one number worth knowing is the half-life — the age at which a match counts for half of a fresh one. Their fitted rate works out at roughly 107 days, about three and a half months.
Two properties follow, and both matter for the argument below. Nothing is ever discarded: old matches keep contributing a sliver, which stabilises teams with thin recent records. And nothing old contributes much: at that half-life a result from two years ago carries under one percent of last weekend’s weight. The full weight-by-age tables, and what each choice of half-life costs in sample size, are in time decay in football prediction models.
Why five matches is the worst of both worlds
Now compare that curve with the form table’s implicit rule: the last five matches count fully, everything before them counts zero.
Take a side playing roughly weekly. Under a 107-day half-life, the total weight across its whole history sums to about 22 matches’ worth, but the effective sample size — the count that actually governs how noisy the estimate is — works out near 44 matches. Roughly a season’s worth of information, concentrated on the recent end.
A five-match form table has an effective sample size of five.
That gap decides everything, because football is a low-scoring sport and goal counts are noisy. Suppose a team’s true rate is 1.50 goals per match. Over five matches it expects 7.5 goals, and Poisson variation alone gives a standard deviation of about 2.7 — so the observed rate lands roughly 1.50 ± 0.55, an error of ±37% before anything real has changed. Over 44 effective matches the same arithmetic gives ±0.19, about ±12%.
Read a five-match window as evidence of a change in ability and you will be reading noise most of the time. The streaks look meaningful precisely because they are common. A constant-strength team with a genuine 45% win probability wins five in a row with probability 0.45⁵ ≈ 1.9%; across twenty clubs and the thirty-odd overlapping five-match windows in a 38-round season, that is on the order of a dozen five-win runs per season from teams whose strength never moved at all. Windows overlap, so this counts runs in progress rather than distinct streaks — but the point stands. Hot streaks are what constant ability looks like when you sample it five matches at a time.
Raw form ignores who the opponent was
The second failure of the coloured letters is that they are unadjusted. Five wins against the bottom half and five wins against the top half produce the same W W W W W and mean entirely different things.
A rating model gets this right by construction, because it never scores results in isolation. Beating a strong side moves ratings more than beating a weak one; losing narrowly away to a leader can raise a rating rather than lower it. The output that looks like “form” is really performance relative to expectation, weighted by recency — a quantity a form table cannot represent, since it has no column for who was on the other side.
Goals-based models add a third refinement: they read margins, not just outcomes. A team grinding out four 1–0 wins and a team winning four matches by three goals share a form line and do not share a future. This is the same discipline that applies to head-to-head records, where a small pile of vivid data invites more weight than it can carry — the reasoning is set out in head-to-head records in football predictions.
So how many matches should you actually look at?
The honest answer is that the question is malformed, and the reason is worth stating plainly: there is no window. A well-fitted model does not look at the last five matches, or the last ten, or the last twenty. It looks at all of them, at descending weight, and the thing you are really choosing is how fast that weight falls away — a single parameter, tuned on matches the model never saw. How that parameter is chosen, and how flat the optimum usually is, is a subject of its own.
What that leaves for a reader is a better question than “how many matches”. Given any form argument, ask instead:
- Over what horizon? A run of five is a claim about roughly five weeks. Under a three-and-a-half-month half-life, five weeks of results are worth perhaps a fifth of the evidence already in a team’s rating.
- Against whom? Unadjusted results cannot distinguish five wins over the bottom half from five over the top.
- Did the underlying rates move, or only the outcomes? Four 1–0 wins and four 3–0 wins produce the same form line and imply different teams.
A run that survives all three questions is genuinely informative. Most do not get past the first.
What form still cannot buy you
Even a well-tuned decay curve leaves two limits standing.
Form is the most public information in the sport. It is printed next to the fixture. Whatever a run of results implies about a team’s strength has been visible to every market participant since the final whistle, so it is in the price already. An edge would require reading form better than the consensus does — usually by opponent-adjusting it, which is exactly what the crowd’s casual version fails to do.
Runs regress. A team performing above its rating tends to drift back toward it, because part of every run is variance and variance does not persist. Extrapolating a streak forward is the single most common error in form-based reasoning, and it is systematically wrong rather than randomly wrong.
The FootInsights model reads years of results, fixtures and prices; how it turns those into a probability is described on the how our predictions work page. Whether it weighs form well is not something we ask readers to take on faith: every settled prediction, with its probability and its outcome, is published on the track record, which is where over-reacting to a run of results would eventually show up as poor calibration.
Which is why, when a preview leans its whole case on a row of coloured letters, the letters are usually the weakest evidence on the page — and always the most widely seen.