How the values are built

For every graduate we walk down this ladder and stop at the first rung that holds. The badge next to his value tells you which one it was, on every single card — a number without its provenance is not published here.

It is a gradient-boosted decision tree ensemble — a few hundred small trees, each one correcting the last one’s mistakes. It is not a neural network, it does not read text, and it has no opinions. It has only ever been shown one kind of thing: transfer fees that were really paid and publicly disclosed, parsed from Wikipedia’s per-window transfer lists, together with the buying club, the date, and the player’s record.

Old fees are restated in today’s money first. The correction is not an economic index we found somewhere — it is derived from our own transfer table, as the median disclosed fee per twelve-month window, smoothed, with the most recent window set to 1.0. Football inflation measured with football fees.

The model predicts the logarithm of the fee, not the fee, so it is fitted on proportional error: being €4m out on a €10m player is the same size of mistake as being €40m out on a €100m one. The result is then clamped between €0.5m and €222m, the all-time record fee — a guardrail, not a judgement. One editorial constant sits on the end: a market-level multiplier, currently 1, which exists so the whole output can be re-levelled against published press aggregates if it drifts. At 1 it does nothing at all.

This is the complete list. There is no scout report in here, no hidden adjustment, and no field where anybody types a number they like. If a fact about a player is not on this list, it did not move his value.

It cannot price a player nobody has sold. Decision trees do not extrapolate. Whatever the largest fee in their training data is, that is the highest number they can produce, for anyone, ever. Ours is €146.3m, and only 9 disclosed fees in the whole corpus clear €80m. So the very best players — the ones whose clubs would never sell them, who therefore generate no fee for anyone to learn from — come out systematically low. This is not a bug we are about to fix; it is what a fee-trained estimator is. It is also the entire reason rung 1 exists: when the press reports a rejected bid or a release clause, we use that instead and show you the article.

It is often wrong, and we publish by how much. Every run trains on the older fees, then predicts the last 120 days of real disclosed fees it has never seen — 50 of them. The average miss is €18.4m and the median miss €10.2m, the gap between them being the handful of big transfers it gets badly wrong. Treat every EST on this site as a number with that much air around it.

Its view of the market is English-language. Fees are parsed from Wikipedia’s “List of English football transfers” pages, because those are the ones that carry machine- readable amounts. Deals between two Serie A clubs, or two Brazilian ones, are largely invisible to it — so it has learned what the Premier League pays more thoroughly than what anyone else pays.

It reads today’s record on yesterday’s transfer. Appearances, goals, honours and readership are current totals, used even for a fee paid three years ago — no historical snapshot of those numbers exists in open data. Club strength is likewise today’s Elo table, not the table on the day of the deal. Both make the model slightly too clever about the past, and we would rather say so than bury it.

It has never seen a fee that was not disclosed. Undisclosed fees, add-ons, sell-on clauses and swap deals are all outside its world. The number it produces is an estimate of a headline fee, not of a club’s accounting.

When we get one badly wrong and the player then moves, the miss goes in the ledger with the real fee next to it. The track record keeps the bad calls as carefully as the good ones.

The valuation pipeline is Python, SQL and scikit-learn. There is no model of language in it, no API call to one, and no place where free text becomes a euro amount. Run it twice on the same snapshot and you get the same numbers to the cent: the training data is fixed by the snapshot date, the algorithm is deterministic, and the one random element — the seed the trees are built with — is pinned to a constant in the source. The estimate on a card is a pure function of the published data.

A language model does appear in this project, once, in a role with no write access to anything: as an auditor. It reads a published value, searches the web, and reports whether a named outlet has published a relevant FACT — a bid, a clause, a club statement, a fee. It returns flags and URLs. It is explicitly forbidden from supplying a valuation of its own, and if the only figure it can find traces back to a valuation aggregator, the required answer is “no fact found” and our number stands unchanged. Anything it flags is read by a human, who writes the anchor by hand with the source URL and an expiry date, or does not.

An LLM that “recalls” a player’s market value is quoting an aggregator from its training data. That is why it is never asked. No value from Transfermarkt or any other valuation database is scraped, stored, cited or laundered through anything in this project; the whole dataset is built from Wikidata (CC0), Wikipedia (CC BY-SA), ClubElo and named press articles.

You do not have to take any of this on faith — the code is public: the model, the valuation ladder, every anchor with its source and the auditor’s rules.

Open an issue. Tell us the player and what you think is wrong, and if you have one, attach the thing that settles it: a link to an article reporting a bid, a clause, a completed transfer or a club statement, with its date. That is what we can act on — a reported fact becomes an anchor at the next refresh, and the card starts showing your source.

What we cannot act on is a figure from a valuation site, however confident. It is not a fact about the market, and it is not something we are able to cite. “He is obviously worth more than that” is welcome too, mind — it is often right, and if it is right about several players at once it is a model problem rather than a data problem, which is more useful to us than any single correction.

Open an issue · corrections ship with the next refresh, each one recorded in a public file with its reason. Wrong positions, missing graduates and eligibility arguments go to the same place — the rules cover those.