Player Rating

Oraca puts every player in its database on a single, position-aware, cross-league ability scale — so a centre-back in the fourth tier and a centre-back in a major league can be ranked on the same axis — and adds a second, independent signal: undervaluation, the part of a player's ability that his club context does not explain. The first answers how good is he? The second answers is the market sleeping on him?

In one line: there is no honest single "player rating" that spans positions — a cross-position number either collapses into a league table of good players on good teams, or buries every attacker under every defender. So we rank players within their position on a globally comparable ability score built from three independent signals, and we separately flag the ones whose ability runs ahead of the club they play for.

Why it is built this way

The obvious request — "give every player one number" — is structurally impossible without ball-by-ball tracking data, and we tested the shortcuts. Each one failed a leak-free, out-of-sample test, and the failures define the design:

• A single cross-position number collapses toward a club table — it ranks good players on good teams — and systematically buries attackers, whose contribution to goal difference is noisier than a defender's. Position is not a nuisance to normalise away; it is the unit of comparison.
• Folding club strength into the score simply re-ranks players by who they already play for. It launders the market's existing consensus back in and destroys the one thing a recruiter actually wants: a player who is better than his shirt. Club strength is kept strictly out of the ability score and used only, separately, to measure undervaluation.
• A dramatic "star" scale that stretches the top end looked impressive and predicted worse out of sample. We kept the plain one.

What survives is a per-position ranking plus a disagreement-with-consensus signal — both validated against what players actually went on to do, not against the eye.

How a player's ability is measured

The ability score combines three independent views of a player. Each measures something the others miss; each is weighted toward recent seasons (older form fades smoothly rather than dropping off a cliff); each is put on a common scale and then added together. We tried learning a clever per-position weighting of the three — a plain sum predicted just as well out of sample, so we kept the simpler, more honest thing.

1 — Opponent-adjusted production. What a player does on the pitch, ranked against his own kind and adjusted for who he did it against. A striker's finishing is compared to other strikers that season, not to full-backs; a defender's actions to other defenders. Then every match is scaled by the strength of the opposition — a strong performance against a strong side counts for more than the same numbers against a weak one. Because the opponent-strength rating underneath this is calibrated across borders, this is the step that makes production comparable between leagues rather than only within one.

2 — Impact on the scoreline. Production statistics reward what's visible. This signal captures the rest: what happens to the score while a player is on the pitch, with the quality of his teammates and opponents mathematically partialled out. The naive version of this is hopelessly noisy — players share minutes, so their individual effects blur together. The fix, drawn from a long line of basketball-then-football research, is to shrink each player's estimate toward an informed expectation — his own production — rather than toward the average. A player with thin or unlucky minutes is pulled back toward what his output says he is, not toward zero. Here too, no club strength is ever added in.

3 — Ceiling games. Averages hide ceilings. A quietly steady player and a player capable of taking a game over can share the same mean; they differ in the tail. This signal looks at a player's best performances — the top end of his match-by-match distribution — to capture the upside an average erases.

The three are summed into one ability number, and the figure a recruiter sees is its percentile within the player's position group (goalkeeper, centre-back, full-back, holding midfielder, central midfielder, attacking midfielder, winger, striker). The scale is anchored on regulars — players with enough recent minutes to trust — so that fringe and small-sample players are placed on the same axis without distorting it.

The undervaluation signal

The second output is the one a recruiter cares about most: how much better is this player than his situation would predict? It is the part of a player's ability that a small model of the market consensus — his club's strength, his league's tier, and his age — cannot explain.

A positive value means a player is more able than his club, his division and his age would lead the market to expect. That is the edge: a disagreement with the consensus that the club and the league already encode. This is precisely why club strength is excluded from the ability score in the first place — if it were baked in, undervaluation would be near-zero by construction, and the signal would tell you nothing you didn't already know.

How we keep it honest

The signal was validated leak-free. We trained on older seasons, then tested on a later season the model had never seen, asking it to predict whether a player would subsequently move up — to a stronger club or a higher tier.

• The undervaluation signal beats the obvious baseline — the club's own strength — at predicting an individual move up. It carries information the league table does not.
• Among under-26 regulars, the players in the top 5% by undervaluation went on to move up at roughly 1.3× the base rate, by a margin that is statistically real.
• Face validity holds: the top of each position is sane (the players you would name); average ability falls steadily as league tier drops — the cross-league property we want — and the overall top 100 is not just a list of one or two big clubs' squads.

Past model performance is not a guarantee of future results — but the signal was checked the hard way, against outcomes it had never seen, before it was allowed near a recommendation.

What it can't do

We would rather tell you the edges of the tool than pretend it has none:

• Production, not tracking. Without ball-tracking data, off-ball value — pressing, positioning, the run that drags a defender out of position — is only partly captured. The position cohorts and the opponent adjustment reduce role and style leakage; they do not erase it.
• Goalkeepers are handled separately. Undervaluation is confounded by workload for keepers — a keeper behind a leaky defence faces more shots, which can flatter his numbers — so they need, and get, a keeper-specific treatment rather than the outfield one.
• Coverage, not model. The rating only sees the leagues we cover. Genuinely uncovered competitions are blind spots — a question of data reach, not of the model.
• Undervaluation is deliberately coarse. It is a simple residual against three controls, kept simple on purpose so that it can never quietly relearn the ability score it is supposed to be independent of.

The research behind it

The rating draws on the plus-minus literature (basketball first, then football) and on classical statistical shrinkage; the opponent-adjusted production composite and the ceiling term are Oraca's own. The undervaluation idea borrows the framing of market mispricing from sports economics.

• Rosenbaum, D. T. (2004). Measuring How NBA Players Help Their Teams Win. — The origin of Adjusted Plus-Minus: isolate a player's effect on the scoreline by controlling for everyone else on the pitch.
• Sill, J. (2010). Improved NBA Adjusted +/− Using Regularization and Out-of-Sample Testing. MIT Sloan Sports Analytics Conference. — Regularised plus-minus with explicit out-of-sample validation; why our impact signal is regularised rather than raw.
• Kharrat, T., McHale, I. G. & Peña, J. L. (2020). Plus–minus player ratings for soccer. European Journal of Operational Research. — Adapts regularised plus-minus to football's low-scoring, substitution-heavy structure.
• Matano, F. et al. Augmenting adjusted plus-minus in soccer. Journal of Quantitative Analysis in Sports. — Shrink each player's plus-minus toward an informed prior rather than toward zero. This is the shape of our impact signal, with opponent-adjusted production as the prior.
• Hoerl, A. E. & Kennard, R. W. (1970). Ridge regression: biased estimation for nonorthogonal problems. Technometrics. — The shrinkage that tames the blur between players who share minutes.
• Hakes, J. K. & Sauer, R. D. (2006). An economic evaluation of the Moneyball hypothesis. Journal of Economic Perspectives. — The market systematically misprices particular skills; the gap between objective production and market valuation is the exploitable signal. The framing behind the undervaluation residual.