Niana Analytics · football data hubMetrics · tactical data · formats · video analysis

Home › Metrics

Why per-90 numbers lie on small samples

Niana Analytics desk · Metrics ·
Why per-90 numbers lie on small samples

Per-90 numbers are the standard way to compare players who played different amounts. They are also the fastest way to produce a confident conclusion from four matches. This guide takes the arithmetic apart: what a per-90 value really is, how much a single event moves it, and at what point the number deserves to be compared with anybody else's.

In this guide (7)
  1. The question
  2. What per 90 actually is
  3. The arithmetic of a single event
  4. How much noise is inside the number
  5. Why thresholds like "900 minutes" exist
  6. What to do instead
  7. Where this article stops

The question

A player plays 180 minutes and scores once. Another plays 1 800 minutes and scores ten times. Both have a per-90 rate of 0.5. Are those two numbers the same measurement? They are computed the same way, but they do not carry the same information, and the difference is not a detail — it is the whole problem.

The question this article answers is narrow and practical: given a count of events and a number of minutes, how much of the resulting per-90 figure is signal and how much is the arithmetic of a small denominator.

What per 90 actually is

A per-90 rate is a proportion rescaled to a fixed exposure: events divided by minutes, multiplied by ninety. Nothing more. It does not know how those minutes were spent — starter or substitute, winning 3-0 or chasing a game, one position or three — and it does not know how many events went into the numerator.

That last point matters most. A per-90 figure throws away the sample size, because two very different inputs produce the same output: one event in 180 minutes and ten events in 1 800 minutes both return 0.5. Anyone reading the output alone cannot tell which one they are looking at.

The arithmetic of a single event

The table below is pure arithmetic, not a claim about any player. It shows one, two, three and six events across different amounts of playing time, with the per-90 rate each combination produces.

Events90 min180 min360 min900 min1 800 min
11.000.500.250.100.05
22.001.000.500.200.10
33.001.500.750.300.15
66.003.001.500.600.30

Rates per 90 minutes. Read a row: the same number of events produces a rate that falls fourfold when the minutes quadruple.

Two conclusions follow immediately. First, a per-90 rate of 1.0 built on 90 minutes and a rate of 1.0 built on 1 800 minutes are not comparable numbers; they are the same formula applied to samples that differ by a factor of twenty. Second, near the top of the table a single event moves the rate by a whole unit, which means one deflected shot or one accidental assist can separate two players by more than a season of work at the bottom of the table.

How much noise is inside the number

Counting events in football behaves, to a rough first approximation, like a Poisson process: a rare event arriving at a steady underlying rate. That approximation lets us put an interval around a count instead of pretending it is exact. For a count of k events, a 95% interval is approximately k ± 1.96 × √k.

Events counted95% interval around the countRelative width of the interval
30 to 6.4±106%
50.6 to 9.4±88%
103.8 to 16.2±62%
2011.2 to 28.8±44%

Illustration using the Poisson approximation on the count itself, before rescaling to 90 minutes. Real event counts are not strictly Poisson, but the order of magnitude is the point.

The message is uncomfortable for anyone who compares per-90 tables: at three events the uncertainty is wider than the value itself. At twenty events it is still nearly half the value. A player with a per-90 of 0.30 on eight events and a player with 0.25 on forty events are not ordered by talent; they are ordered by how much we happen to know about each of them.

Why thresholds like "900 minutes" exist

Analysts often refuse to publish per-90 rates below a threshold — 450, 600 or 900 minutes, depending on the desk. The number is arbitrary, but the instinct is right: it is a crude way of demanding that the numerator be large enough for the ratio to mean something. A threshold does not fix the problem, it just keeps the worst cases out of the table.

A better habit is to publish the count and the minutes next to the rate. "0.5 per 90 (1 in 180)" is a sentence a reader can judge; "0.5 per 90" is a number that hides its own weakness.

What to do instead

  • Show the raw counts. Nine goals in 1 100 minutes says more than 0.74 per 90, because the reader can see how thin the sample is.
  • Compare like with like. Minutes as a starter and minutes as a substitute are different jobs; so are minutes in a dominant team and minutes in a relegation side.
  • Shrink towards a prior. When samples are small, the honest estimate is pulled towards the positional average rather than reported raw. A winger with two goals in 200 minutes is not a 0.9-per-90 scorer; he is probably an average winger with a very short record.
  • Ask what decision the number feeds. For a scouting shortlist, a wide interval is acceptable if it only triggers a second look. For a published ranking it is not.

Where this article stops

The Poisson framing assumes a constant underlying rate, and football does not work that way: opponents differ, roles change, a player returns from injury, a team changes manager. The interval in the table is therefore an illustration of scale, not a rigorous confidence statement about any real player. It is also why the honest conclusion is not "compute the interval" but "do not compare per-90 numbers built on small samples at all".

For what a rate per 90 means in a reading context, see the metrics section. For the same problem applied to a whole match — where the sample is the period, not the season — see the field tilt case study. Definitions of the terms used here are in the glossary.

About our numbers. Every figure on this site describes a method, a formula or a competition rule. We deliberately do not publish invented player statistics, match scores or rankings, and we add no links to outside sites. Read every metric as a tool with a definition, not as a verdict.

Read next

Metrics

What is xG and PPDA? A football metrics primer

Expected goals and passes allowed per defensive action explained from first principles: what each number measures and how to read it.

Metrics

Expected goals (xG): how the metric is built

Shot location, angle, body part and defenders: the ingredients of an xG model and why a single number never tells the whole story.

Metrics

Expected assists (xA): valuing the final pass

How xA scores a pass by the quality of the chance it creates, and how to use it without overrating a creative midfielder.