PitchMetric

We spent this research cycle finding out, using real historical data: about 187,000 valuation snapshots across roughly 9,000 players in the top five European leagues, going back to the mid-2000s.

Three different ways to forecast a price

We tested three well-established forecasting philosophies against each other, honestly — meaning we never let a model see the future, and we tested it both on players it already knew about (forward in time) and on players it had never seen at all.

• Time-series forecasting. Treat a player's value history like a stock chart and extrapolate the trend. Intuitive — and it turned out to add no real predictive value at all, however carefully we tuned it. It badly underreacts when a player's trajectory bends, and wasn't even better than simply assuming a player's price won't change. We dug into exactly why (a real bug in our first attempt, then a real ceiling even once that was fixed) rather than quietly dropping it — more on that below.
• Momentum. Rather than looking at one player in isolation, pool the patterns across every player: is this player's price rising fast, and is their output (goals, assists, minutes) rising too? This approach doesn't need years of history on a single player — it borrows patterns from everyone else's careers. This turned out to be the strongest of the three individual approaches.
• Bayesian updating. Start with an informed opinion of what a player should be worth — anchored on their last known transfer fee, or on typical value curves for their age and league if there's no recent deal — and then update that opinion using their recent on-pitch output. The appeal isn't just the prediction itself, it's that this method also tells you how confident to be in it.

On top of these three, we built a fourth, blended model that combines the two strongest approaches (momentum and Bayesian updating — full details of exactly how deliberately kept behind the curtain for now) — and it beat both of its ingredients on our held-out tests, not just the weaker of the two.

The honest part: it's a lot easier for players you already track

One of the more useful things this research surfaced wasn't a number — it was a pattern. Prediction accuracy depends enormously on how much history you have on a player. For players with a rich track record, our best model's typical error was in the range of 14–15% of their value. For the full universe of players, including fringe and youth players with thin histories, that error is meaningfully higher.

We also tested the models against players they'd genuinely never seen before (not just future dates, but entirely new careers) — and here the honest picture is more encouraging than we expected going in: the models still clearly beat the simplest possible baseline (assume the price doesn't change), especially at correctly calling which direction an unfamiliar player's value is about to move. What doesn't hold up for unseen players is our blended model's edge — it's tuned for players with a track record, and a plain momentum read does just as well or better once a player is a genuine unknown. The lesson: these tools are strongest as a way to track players you're already watching forward in time — less a crystal ball for scouting a name you've never heard of, more a sharper lens on names you already follow.

What did fixing our mistake actually buy us?

Worth being honest about, since it's the most useful thing that happened in this research cycle: our first pass at the time-series approach had a real bug in how it was being tested — it made the approach look usable when it quietly wasn't. Once we caught that and re-tested everything properly, it turned out time series wasn't pulling its weight at all, however we tuned it. The honest fix was to drop it from the blend entirely and lean harder on the two approaches that actually did the work — momentum and Bayesian updating — and to sharpen the pattern-detection underneath both of them.

That's not a rounding error — it's roughly a 13% reduction in typical prediction error, and it came entirely from being willing to distrust a result that looked fine on the surface and dig into why. We'd rather publish the slower, more honest version of this research than the faster, flattering one.

What actually moves the needle — and what it means for your squad

Here's the part that matters most if you're playing FPL or La Liga fantasy rather than running a scouting department. We cross-referenced the features that predict market value against real Fantasy Premier League scoring data for hundreds of Premier League players. The question: does whatever predicts a rising price tag also predict fantasy points?

The single strongest signal for fantasy output wasn't market value at all — it was recent output itself. A player's trailing goal-and-assist rate correlated with their FPL points-per-90 roughly 60% more strongly than their current market value did. In plain terms: the market price is a noisier read on fantasy potential than just looking directly at what a player has actually been doing on the pitch. Value bakes in reputation, potential, and club prestige — none of which score you FPL points.

Does a player's own fantasy price predict their transfer value?

One more thing we tried, since it's a natural question once you're staring at both datasets side by side: FPL runs its own weekly-updated "valuation" of every player — its price. That's a genuinely independent opinion of a player's worth, from a different institution, updated by a different mechanism than Transfermarkt's own valuation. Does knowing a player's FPL price, or how many fantasy points they've racked up recently, help predict where their transfer value is headed?

Short answer: not really, and the reason is a nice bit of intuition once you see it. FPL price is deliberately compressed for game balance — two squad players can cost the same in FPL while being worth wildly different amounts on the transfer market. It just isn't fine-grained enough to track real value differences. Recent fantasy points fared a little better on their own, but added almost nothing once our model already knew a player's actual output (goals, assists, minutes) — it was measuring mostly the same thing through a different lens. A tidy negative result, but a real one, and exactly the kind of thing worth checking rather than assuming.

What's next

This is the first entry in our new Deep Research series — periodic, honest, methodologically serious dives into questions we're genuinely curious about, using real data and real held-out testing rather than vibes. Part of that honesty: an earlier pass at this analysis had a real mistake in how the time-series approach was being evaluated, and fixing it (rather than quietly deleting the disappointing result) is exactly what led us to the stronger blended model above. We'll keep that habit up — next on the list is validating our blend against players our models have never touched at all, and seeing whether a smarter time-series approach can add anything once it's given a fairer shot.

If you want the numbers behind your own squad, head to Transfer Finder or Predictions — both are built on the same “trust the underlying output” philosophy this research validated.