Fair to the referees, and fair to the statistics.

Where the data comes from

Every grade is built on the Premier League's independent Key Match Incidents (KMI) Panel — a five-person body of former players, former coaches, and one representative each from the league and the referees' body (Pro Ref). It reviews a selected set of decisions — the contested ones — and rules whether the on-field call and the VAR call were each correct.

This matters because it means we are not substituting our own opinion for the referee's. We don't count the decisions fans were angry about. We count only the ones an independent panel formally ruled were wrong. If an incident isn't on the panel's list, we treat it as correct — even if it was controversial.

Absence from the panel's list is not a clean sheet. Because the panel looks at a selected set, a decision it never reviewed is one we know nothing about — not one it approved. That cuts both ways and we hold to it in both directions: we never count an unreviewed decision as an error, and we never present "no decisions reviewed" as a record without a blemish.

How we see the panel's verdicts. The KMI Panel doesn't publish its rulings in a public data feed. They reach the public through Dale Johnson — now BBC Sport's refereeing and VAR correspondent, previously at ESPN — whose week-by-week Key Match Incidents reviews are our source: ESPN's panel write-ups for 2023–24 and 2024–25, and Dale Johnson's BBC Sport VAR review for 2025–26. Every incident we grade is traceable to one of those published reviews, including the on-field and VAR vote split where the panel gave one.

From 2026–27 the League publishes the docket itself. The Premier League now puts out the panel's complete round-by-round results — every reviewed decision with both votes, the lead official, and the panel's reasoning in its own words. That is a different kind of source from the three seasons before it, and the difference matters when you read a figure that spans them:

• It carries correct decisions, not only errors. The relayed seasons catalogue what went wrong; the League's own document lists what the panel looked at and upheld as well. So the same referee can appear to have "more incidents" from 2026–27 onward purely because more of his work is now visible.
• It names the lead official. Each row says whether the decision was the referee's or an assistant's. An assistant's error counts for the match and the clubs and is never charged to the referee — a distinction the earlier seasons could not make.
• We do not pool the two. They have different inclusion rules, so combining them would invent a denominator. Season figures stay season-scoped, and anything spanning the boundary says so.

How a grade is built

The final number is what's left: how a referee did relative to a typical official over the same number of matches. Negative is better than the league average; positive is worse; zero is exactly average.

Why there's no difficulty adjustment

It's tempting to grade referees against the difficulty of their fixtures — surely derbies and top-six clashes breed more mistakes. We tested that, hard, and it isn't true. Two independent checks agreed:

• A difficulty model built from opponent strength and match stakes, tested out-of-sample — fit on two seasons, used to predict the third — correlated with actual errors at just +0.09, indistinguishable from zero.
• A full model on all 1,140 matches, using only things known before kickoff — team quality and league position, each side's disciplinary record under other referees, derbies, matchweek, relegation and title stakes — did no better than assuming every match is equally likely to produce an error.

The finding is worth stating plainly: refereeing errors are not predictable from the character of a match. Big games and derbies are no more error-prone than a mid-table Tuesday night. So we don't pretend otherwise — every referee is measured against the same flat league average, and we'd rather say that than dress a flat number up as a difficulty model it isn't.

What the league trusts him with

Separately from the grade, each referee's profile shows the kind of fixtures Pro Ref assigns him — a marker of standing and trust, not a measure of accuracy. Unlike errors, this does separate referees sharply: some are handed the marquee games week after week, others rarely. We count, pooled across seasons, how many of a referee's matches are:

• Two-big-six clashes — both clubs from a fixed list of six.
• Derbies — a fixed list of twelve recognised rivalries.
• High-stakes run-in fixtures — matchweek 33 or later with a side in the top seven or bottom six by the table as of that round (using only earlier matchweeks — again, no hindsight).

Both lists are fixed in advance, which is the point: they carry no hindsight from where teams happened to finish, so a club having a good season cannot retrospectively turn its fixtures into marquee ones.

A “marquee” match is any of the three. We report the share, the counts, and the single biggest fixture a referee took. It reads fairly at both ends: a light assignment load is a fact about the league's choices, not a mark against the official.

What we found — and won't hide

The honest headline is that Premier League referees are far more alike than the weekly outrage suggests. Most sit within a few thousandths of the average, well inside the margin of error. The real differences live at the very top and very bottom — and even those are smaller than you'd guess.

How we name an official — and when we won't

Naming a referee is easy: the match record says who refereed. Naming the VAR is not. Every VAR name we hold comes from a single secondary source, and one source is enough to credit an official but not to accuse one. So a name reaches an error surface only when a second, primary source agrees.

Every appointment is therefore in one of three states, and they are not interchangeable:

• Verified — premierleague.com's own match-officials listing names the same person. The official may be named beside the error.
• Unchecked — never confirmed, with nothing against it. Counts in the totals, with the coverage stated, but any surface naming a person says "the VAR" instead.
• Contradicted — the primary source names a different official. The error attributes to nobody.

The principle underneath, which we apply in both directions: absence of confirmation is not evidence of error, but a primary source naming someone else is. Collapsing those two would either paralyse the record — treat every unchecked appointment as unusable — or launder a known-bad one into an accusation.

A contradicted error is never deleted. It stays in the published total and moves to an unattributed bucket. Dropping it would shrink the error record to make a sourcing problem disappear, which is the failure this rule exists to prevent. Of 72 VAR-layer errors, 54 sit on a verified appointment, 1 is unchecked, 1 is contradicted and attributed to no one, and 16 have no appointment on record at all.

One further limit, and it is the League's rather than ours: where an incident was an assistant referee's call, the published record names the role and never the person. We hold both assistants for every fixture and still do not name one, because nothing says which of the two raised the flag. Those incidents count for the match and the clubs, and are never charged to the referee.

Net xG Impact — what an error was worth, and which way it went

A confirmed error is a count. Net xG Impact says how much it was worth, in the one currency football already has: expected goals.

The sign, because three pages show one and it decides how you read them. A club's net xG is what errors gave it, minus what they took away:

• Positive — errors fell in this club's favour on balance. It was helped.
• Negative — errors went against it on balance. It was hurt.
• Zero — the two sides cancelled, which is the ordinary case, not a sign that nothing happened.

Per incident the figure is signed, symmetric and sums to zero: whatever one club lost the other gained, so across the whole league the ledger nets out exactly. That is a property of the definition rather than a finding, and it is why a club with a large number of errors in its matches can still show a net near zero — a reason the site shows the counts for and against beside the net, never instead of it.

Two constants carry every xG figure on this site. Both are stated here because everything else is arithmetic on top of them:

• A penalty is 0.76 xG — the long-run Premier League conversion rate. A penalty wrongly given or wrongly denied moves that much.
• A dismissal is worth 0.012 xG per remaining minute to a home side, 0.017 to an away side — from Titman, Costain, Ridall & Gregory (2015), who measured how a one-man advantage changes each side's scoring rate over the rest of a match. The split is asymmetric because a home red and an away red do not cost the same. A red in the 20th minute is worth more than a penalty; one in stoppage time rounds honestly to nothing.

A goal wrongly allowed or disallowed is not priced at all. Its value is the expected goals of the shot itself, and we hold match totals rather than shot-level data. A constant would invent a number and a zero would say the decision did not matter, so those errors are counted and carry no value.

An error we cannot value is held out of the xG, never entered as zero. A missed dismissal leaves no event in the match feed — a card that was never shown is not recorded — so there is no minute to price it from. Those errors keep their place in every count, and the surfaces that total xG say how many they set aside.

The figure prices the decision, not the consequence. It does not know whether the goal was scored, and it is not a claim that the result would have changed.

The style axes — how a referee runs a game

Style is descriptive. None of it measures accuracy: a referee who whistles often is not worse than one who lets play run, he is refereeing differently. Three axes reach a card, each pooled across a referee's whole career:

• Whistle — fouls given per game. High means he inserts himself often; low means he lets play run.
• Tolerance — how many fouls he allows before a booking. High means play flows; low means he reaches for a card quickly.
• Cards — yellows per game.

Why only three, and how the others were tested. Split a referee's own matches into two random halves at random and compare the two figures: an axis that describes the man scores near 1, an axis that mostly reflects which matches he happened to get scores near 0. That is split-half reliability, and it is the first of two tests every candidate axis had to pass. Measured across all 26 seasons:

There is a clear gap between the three we print and everything else. "Tightens late" had been the lead fact on seven of ten cards in one round, on an axis that barely repeats between two halves of a referee's own matches — so it came off.

Two axes were dropped on the second test, not the first, and the distinction matters. Home/away lean and dissent share both repeat well enough (0.485 and 0.471, above several figures the site once printed). They failed a different question: whether a named referee can be told apart from the league at all. Of twenty referees with enough reasoned cautions to test, one sits above the league's dissent share by more than his own error bar and one below — eighteen are indistinguishable. A trait can be real and still be unprintable beside a man's name, and that is the same rule the rest of the site runs on: counts pass, rates fail.

Why never one season. Over a full career the two halves agree closely; over three seasons they barely agree at all. A three-season window measures the window, not the man, so every style figure runs across all 26 seasons we hold and the ordering on those pages is navigation.

How added time's "expected" is calculated

The added-time page compares each referee's real second-half stoppage time against what a transparent model predicts from the visible stoppages in his matches. Here is that model in full.

Go deeper

For the league-wide breakdown of what kinds of mistakes referees actually make — missed by VAR, wrongly rejected at the monitor, below the clear-and-obvious threshold — and how the totals have moved across three seasons, see the Error Anatomy page.