Getting started

FAnalytics Arena benchmarks football analytics products against ground truth, in public, on a protocol anyone can audit.

What it measures

Capabilities, not companies. A vendor does not get one score; it gets a score per benchmark, and each benchmark asks one question with a checkable answer. A provider strong at valuing transfers and weak at normalising across leagues should read as exactly that, rather than as a single number that hides both.

Who it is for

Vendors get an independent measurement they did not mark themselves, and a place to be compared on the same cases as everyone else.

Clubs and analysts get a basis for choosing a supplier that is not a sales deck — including the uncertainty around each number, which is usually where the real answer lives early in a season.

Researchers and journalists get a methodology that can be reproduced or challenged, because every published score names the dataset versions and the formula that produced it.

How participation works

How a participant answers depends on the task, and each task specifies exactly one way in. Transfers uses submission: you fetch the open cases, send one file of predictions, and are scored weekly. A future task where the prediction window is an hour wide — after team news, before kickoff — will use a live endpoint instead, because only then does the Arena need to control the moment of asking.

For Transfers:

1. Read the submission contract and fetch the open cases.
2. Validate your file against a dry run that stores nothing and tells you, per case, what would be scored and what needs fixing.
3. Submit. A case is answered once and never revised, and the model version is recorded with the answer.
4. Scores publish weekly and are appended, never rewritten, so a participant's trajectory across the season stays readable.

You can join at any point

A benchmark that only accepts entrants in August would be a benchmark almost nobody could enter. Registration is open mid-season, and a participant is offered the cases whose outcomes are still open at the moment it registers.

What arriving late costs is coverage, not accuracy: the cases already under way are not offered, and the coverage figure published beside the score says so. Two participants that entered at different points did not answer the same question, and the board is built to show that rather than to hide it.

Where to go next

Quickstart takes you from nothing to a validated endpoint with curl. The Transfers overview explains what that benchmark asks and why it is shaped the way it is, and the submission contract is the page to live in while implementing.

The evaluation protocol states the rules every benchmark follows, and reading a leaderboard explains what a published board is and is not telling you.

The Reference section holds the shared vocabulary every task is written in terms of: roles, metrics, and a glossary of the terms a board uses.