Transfers

A player just signed for a new club. How will he actually perform there?

Every recruitment product answers this implicitly. This benchmark makes the answer explicit, comparable, and scored in public as the season unfolds.

It covers any move between two clubs. A player arriving from another league and a player moving between two clubs in the same league are both transfer cases.

What a participant predicts

Not a number. A distribution.

For each case, and for each metric in that case's mask, a participant returns a probability distribution over percentile bands of the player's per-match performance at the new club, ranked against his positional peers in the destination league.

A prediction for a striker's goals might read:

Read aloud: in seven matches out of ten he does not score, in a quarter of them he scores once, and once or twice a season he scores twice.

Why a distribution, not a point estimate

A point estimate hides what clubs actually care about. Two strikers with the same expected goals per match are not the same signing: one scores in half his matches, the other is quiet for a month and then scores twice in a game. A distribution says which one you are buying, and it lets a wrong prediction be wrong by a measurable amount rather than simply wrong.

It also makes overconfidence costly. A model that concentrates all its mass on one band and misses is penalised more than one that spread its belief honestly — which is the behaviour you want from something a club is going to act on.

Why percentiles, not raw metrics

Because leagues are not comparable in raw terms and squads are not comparable in minutes. A percentile is defined against the player's own positional peers in the destination competition, so "top decile for passing volume as a Premier League central midfielder" means the same kind of thing as it does in LaLiga, even though the pass counts behind it differ.

Bands are cut at cumulative mass rather than at round numbers. That has a consequence worth knowing: a band's percentile width is its share of peer appearances. A band spanning percentiles 0 to 78 contains 78% of appearances. So the league base rate is readable straight off the grid, which is what makes the scoring floor reachable by everyone.

What is scored

The band grid is derived per competition, role and metric, from the last two completed seasons:

Grids differ by role because roles differ: a striker takes four shots where a centre-back takes one, and a grid shared between them would leave bands nothing can land in.

How it is evaluated

Live, weekly, across the season.

1. Predictions lock when they are submitted. One answer per case, never revised, with the model version recorded alongside it. A case stays open all season, so a participant may join at any point and is scored only on appearances after its own lock.

   One exception exists, and it is recorded rather than arranged. The Arena's own baseline predicts the published band widths and nothing else — no player, no form, no result — so it cannot have seen an outcome, and the season's opening matchweeks would otherwise have no realised distribution to show at all. Its submission carries a scored_from date and the reason for it; the alternative, backdating the lock, would have been indistinguishable from a genuine early entry. No other participant is scored on football played before it answered.

2. Each appearance is banded against the frozen reference pool. Every appearance counts — there is no minutes floor, because playing time is already expressed in the percentile a short appearance lands on.

3. Each week, the realised distribution is rebuilt from every appearance since the start of the season, and each locked prediction is scored against it.

4. A snapshot is published and never rewritten.

The realised distribution is rebuilt in full each week rather than accumulated, so a score is always the whole season to date rather than a running total whose early weeks cannot be re-examined.

Scoring in one line

The ranked probability score between the predicted distribution and the realised one, converted to a skill score against the league base rate.

RPS is order-aware — being one band out costs less than being five bands out — and proper, meaning a participant minimises its expected score by reporting what it actually believes rather than by shading its answer.

Zero means level with the league base rate. Above zero means the participant knew something about this specific player that the league average did not. Below zero is meaningful and is published as such rather than clamped: a confidently wrong model belongs below a shrug.

Three columns are published beside it:

• Coverage — scored cases over scorable cases. Refusals and cases never answered both reduce it, so declining is visible and costed without being scored as a wrong answer.
• Confidence discrimination — whether higher stated confidence went with lower realised error. Its own ranked column, never folded into the skill score.
• The interval — every score carries one, because early in a season the interval is the story.

Leagues covered

What to read next

Submission contract is the page to live in while implementing, and Quickstart gets you to a scored submission fastest.

Scoring has the formulas and the worked examples behind every number above. Reading a leaderboard covers what the published columns mean and the four ways a board like this is most often misread.