Does squad market value predict football results?
Published
Ask whether squad market value predicts football results and you get two confident answers, both half right. One says yes, obviously — the expensive squad wins. The other says no, football is chaos and money buys nothing on the day. The useful answer is narrower than either: crowd-sourced squad valuations do carry genuine information about match outcomes, they carry most of it at exactly the moment a results-based model has least, and they carry a hidden copy of information you may already be feeding your model.
What a squad market value actually is
A club’s squad market value is the sum of per-player valuations produced by a community of contributors — not a transaction price, not an accounting figure, and not something any club publishes. Two properties of that construction matter more than the number itself.
It is forward-looking. A valuation is an opinion about what a player is worth from here on, formed by people who have watched him. It is not a record of what he has already done, which is what a results table is.
It is slow. Valuations are revised in batches, on the order of weeks.
Together those make squad value behave like a prior: a summary of expected strength, formed before the next round of matches is played, and updated lazily afterwards.
Why it predicts anything at all
The mechanism is wisdom of crowds. Thousands of independent estimates, each wrong in its own direction, aggregate into something whose errors partly cancel. And the underlying quantity being estimated — what the transfer market would pay for this player — is itself a market price for talent, and talent is one of the inputs to goals.
This works better than sceptics expect. Published work in sports economics has repeatedly found crowd-sourced valuations competitive with, and in some comparisons better than, official ranking systems as a predictor of international match results — a genuinely surprising finding about amateur aggregation, and the strongest argument for taking squad values seriously.
But note what the mechanism is: value is a shortcut to talent. Talent is not the only input to a result, and it is not even the input most prediction models are short of.
What a rating model already knows
A model built on match results — an Elo-style rating, or a Poisson goal model with per-team attack and defence strengths — estimates each side’s strength from what it actually did on the pitch. Give it a few hundred matches and that estimate is very good. Crucially, it has already absorbed everything the money bought, plus everything the money failed to buy: the coaching, the tactical fit, the injuries actually suffered, the players who cost nothing and are excellent.
So the question worth asking is not “does squad market value predict football results?” — it does — but “does squad market value predict results beyond what a team’s own results already tell you?” That is a far higher bar, and for a settled side deep into a campaign the honest answer is: barely. The information is not absent. It is redundant.
Where the value ratio earns its keep
The exception is the cold start, and it is a large exception. Consider the cases where a results-based rating is thin, stale or simply absent:
- A promoted club. Its record was compiled in a different division against different opposition. A rating carried across divisions is an extrapolation, not a measurement.
- The opening rounds of a campaign. Ratings arrive carrying the previous year’s evidence about a squad since rebuilt, and it takes a run of matches before new evidence outweighs old.
- A club that turned over most of its first team. The badge is continuous; the team is not.
In each case the results-based evidence is weakest precisely where a forward-looking prior is strongest. Our own offline testing pointed that way, which is why we ingest weekly squad-value snapshots at all: the signal concentrates in the early window and fades as results accumulate.
One modelling detail is where most casual analyses go wrong. The useful feature is the log ratio of the two squads’ values, not the difference. Football strength behaves multiplicatively: a €400m squad against a €200m squad is roughly the same mismatch as €200m against €100m, even though the first gap is twice the second in euros. Feed a model raw euro differences and you will teach it that every fixture involving a wealthy club is a rout.
Where squad value misleads
It re-reads results you already have. Valuations rise after players perform well, so a snapshot encodes outcomes that have already happened. A backtest joining historical matches to present-day values leaks the future into the past and will look wonderful for reasons unrelated to prediction. Join values as of the match date.
It is a stock, not a flow. The total counts a squad; a match is played by eleven. An injured €80m striker sits in the number and not on the pitch, and a deep, expensive bench inflates a total well past what it contributes on the day.
Age and contract length distort it. A twenty-year-old prospect on a long contract can be valued above a thirty-one-year-old who is straightforwardly the better footballer today. The valuation prices a future asset; a model prices ninety minutes.
It does not travel between competitions. Values are set in a global transfer market whose price level differs enormously across leagues. Comparing two squads inside the same table is reasonable. Comparing a squad in one country to a squad in another is much shakier, and the euro figures make it look far more precise than it is.
Coverage is uneven. No valuation dataset covers every competition, and second tiers are usually the thinnest. A missing value has to degrade gracefully — a model that quietly substitutes a league average for an unknown club is inventing an opinion it does not have.
How to test the claim yourself
If you want a real answer rather than an argument, the experiment is small enough to run in an afternoon:
- Take several seasons of finished matches and join each to the squad-value snapshot dated before kick-off.
- Build a single feature: the log of home squad value divided by away squad value.
- Fit a simple match-result model on the earlier seasons and predict a later, untouched one.
- Score with a Brier score, not accuracy. A hit rate cannot tell you whether the probabilities are any good.
- Compare against two baselines: a results-only rating, and the bookmaker’s closing price. The first says whether value adds anything; the second says whether any of it is news to the market.
Then run the combination — rating plus value ratio — and see whether the pair beats the rating alone, and where. An improvement spread evenly across the whole season is suspicious; one concentrated in the opening rounds and in fixtures involving promoted sides matches the mechanism.
And if a value-only model beats a results-based rating across a full season, go back and check step 1 before celebrating. That is the classic signature of a leaked snapshot.
What we do with it
We take weekly squad-value snapshots and show them as context: on team pages, in league tables such as the Premier League, and in the side-by-side comparison on a prediction page. They are labelled community estimates, because that is what they are.
What they are not is a hidden input. The probabilities we publish come from match results and the goal model; squad value sits alongside them as something you can look at, not something folded silently into a number. If that ever changes, it changes as a new model version, and the public track record keeps every prediction made under every version — the old rows are never edited or removed. A record that can be quietly rewritten when the method changes is not a record.
Real information, already mostly counted
Squad market value predicts football results in roughly the way a regional weather forecast predicts the weather on your street. It is real information, it is coarse, and it is mostly already contained in the more specific measurement you have — what the team has actually been doing.
Its worth to a model is concentrated in the gaps: the promoted side, the rebuilt squad, the first weeks before evidence accumulates. Those are the same gaps where most published predictions are least reliable and least willing to say so.