Who plays like him?

Profile. Nine metrics: xA, key passes, xGChain, xGBuildup, non-penalty xG, shots per 90, headed share, in-box share and xG per shot — adjusted for opponent, venue and league environment (not own-team strength), shrunk toward the role average, then z-scored within the role.

Roles. Seven, from the data: Understat's granular positions clustered on per-90 output. The labels are the research's own. Candidates are compared only with players in the same role.

Ranking. Cosine similarity and Euclidean distance under four profile windows (one, two and three seasons, and the latest 2,500 minutes) — eight checks. A match is robust when it is in the top 10 and closer than 90% of the pool under at least two-thirds of them. Results list robust matches first, then by the median similarity score in your chosen window.

Profile similarity = closer to the player than that share of the destination pool (0–100). It is a rank, not a probability and not a rating.

Match language uses the Study 4 case study's thresholds: a metric gap of ≤ 0.5 SD is a strong match, ≤ 1.0 SD moderate, larger a divergence; a group is “closest” at a mean gap ≤ 0.5 SD and “higher/lower” beyond 0.75 SD.

Precomputed. Every player × league × window was run through the research engine (src/study4/similarity.py Engine, windows w3, w2, recent2500, w1 × cosine, Euclidean); this page only looks results up. Profiles as of 2025/26; pools need at least 900 minutes. Validated by self-retrieval: three-season profiles find the same player in the top 9% of a destination pool after a real league move.