Model watch

New laws, old data: what a rule change does to a prediction model

Published 2026-08-22

Every statistical football model carries an assumption it never states out loud: that the game it was fitted on and the game it is predicting are the same game. This season that is slightly less true than usual, because the Laws of the Game changed over the summer — and a model has no way of knowing it.

This is not a post about whether the new laws are good. It is about a failure mode underneath every prediction site, ours included: when the rules move, the training data quietly becomes a record of a different sport.

What actually changed

The relevant changes for 2026/27, as published by The FA, are a cluster of anti-time-wasting measures:

• Law 15 (throw-in): a five-second visual countdown when a player or team delays a throw-in. If it expires, “the throw-in is awarded to the opposing team”.
• Law 16 (goal kick): the same five-second countdown on goal kicks, with a corner awarded to the opposition if it runs out.
• Law 5 (the referee): a player who receives treatment on the field must leave and remain off it for one minute after play restarts.
• Law 3 (the players): a ten-second limit for a substituted player to leave the pitch, or the replacement waits until the first stoppage after a minute.

These sit alongside the goalkeeper’s eight-second holding limit, sanctioned by a corner to the opposition, and the permission — confirmed in the IFAB’s latest law changes — for competitions to use semi-automated offside technology.

Read as a group they push one way: less dead time, more ball in play — an intervention on the quantity of football played per ninety minutes.

Why this is a different problem from a bad prediction

We wrote recently about what our model knows about promoted clubs, where the failure is missing information about a team. A rule change is a different animal: nothing is missing. Every match in the dataset is complete and correctly recorded. They were simply all played under a slightly different set of incentives, so the relationship the model learned — between team strength and goals scored — was measured in conditions that no longer exactly hold.

Statisticians call this regime change, and it is nastier than missing data for one reason: missing data announces itself. A team with no history is visibly a team with no history. A dataset collected under superseded rules looks exactly like one collected under current rules. There is no null value to check.

Our Dixon-Coles model is fitted on finished results and nothing else. It cannot read a law book. If the new countdowns add even a couple of minutes of playing time per match, the true goal rate ticks up, and a model calibrated on the old rate prices goals markets a fraction too low until new results pull it back.

How big would the effect have to be to notice?

Here is what our dataset actually contains, as of 22 August 2026 — goals per match and the share of matches going over 2.5 goals, by season, across every league we cover:

Look at the spread. Without a single rule change to explain it, goals per match wandered between 2.669 and 2.836 across five seasons — a range of 0.167 goals, or roughly six percent, with 2023 alone sitting 0.15 above the season before it.

That is the noise floor. Any effect from the new laws has to be visible through a background that already moves by that much for reasons nobody fully explains, which means it is detectable only over thousands of matches.

So what does the new season say?

Almost nothing, and that is the point. Of the 17,495 finished matches in our dataset, exactly 187 were played on or after 1 July 2026, when the new laws took effect. Those 187 have averaged 2.797 goals.

It is tempting to line 2.797 up against the 2.745 of everything before it and call it a trend. Don’t. In fewer than two hundred matches, the gap between a quiet weekend and a loud one moves the average more than any rule change plausibly could — and leagues adopt law changes on their own timetables in any case. Anyone claiming in August to have measured the impact of the 2026/27 laws is describing noise with a straight face.

That ratio — 187 against 17,308 — is the whole story. The model’s view of how many goals a football match contains is, for now, entirely a view of the old rules.

What we are doing about it

Nothing clever, which is the correct answer. The fix for regime change in a results-fitted model is the same as the fix for a promoted club: play the matches. Every round adds new-rules data, and because the model weights recent results most heavily, its goal expectations drift toward the new normal on their own — no fudge factor, no thumb on the scale.

What we will not do is guess a correction now and present it as insight. A guessed constant would be indistinguishable from the real effect for at least a season, and it would contaminate the one thing that makes this checkable. Whether our goals numbers run a shade low over the opening months of 2026/27 is a question the public track record will answer in public, whichever way it falls.

The takeaway is not really about our model. Every prediction you read this season — ours, a bookmaker’s, or a site quoting a win rate it will not let you audit — was built on a game that has, in a small and genuinely unmeasured way, just changed. The honest ones will tell you that.