Methodology · like-for-like player map · v2

How the like-for-like map is built — and what I got wrong the first time

This page exists because the first version was built by intuition and presented with more confidence than it had earned. Below: exactly what v1 did, what the published and open-source methods do, what v2 does instead, and the tests that show where it holds and where it does not.

Correction, 28 Aug (later the same day). A per-table minutes check shows the archived FBref release's 2024/25 possession, defence and misc tables cover only ~20% of the season (median 4.7 nineties vs 21.9 in the standard table). Six of the twelve metrics below were therefore ~5× too low for 2024/25. Everything in §§4–8 built on 2024/25 is superseded by the 2023/24 rerun in §6b and §8b. Understat-based numbers (the register, Marcel) are unaffected.

1. What v1 actually was

28 metrics chosen by me, from FBref's seven stat tables, on the reasoning "these describe how an attacker plays." No template, no literature check, no redundancy check.
PCA was used for the picture only. Similarity was Euclidean distance on all 28 z-scored metrics. That double-counts anything measured twice — npxG, shots and npxG-per-shot are three views of one thing; so are five passing metrics.
No k-means was run. The "archetype quadrants" language came from the public method being discussed; v1 never clustered anything.
No possession adjustment, no team-context residualisation, no validation.

The headline conclusion — Vlahović and Kolo Muani are not similar players — survives every change below. The neighbour lists do not. That distinction is the point of this page.

2. What already exists, and what was taken from each

Source	Method	What v2 takes
StatsBomb radars (Knutson, 2014→; the position templates the industry copied)	11 fixed metrics per position, per 90, axes scaled to the 5th–95th percentile of the position, defensive metrics possession-adjusted	The metric list. Striker: xG, shots, touches in box, shot-touch %, xG assisted, pressures, pressure regains, aerial wins, turnovers, successful dribbles, xG/shot. Wide: + box-cross %, fouls won
Soccermatics (Sumpter, open course)	Count actions with context filters, per 90 or per possession-minute, convert to percentile within position and league, compare like with like	Position-relative comparison; the warning that tackles and pass completion are context, not quality
PlayeRank (Pappalardo et al., ACM TIST 2019)	76 event features; weights learned by a linear SVM against match outcome (AUC 0.89); 8 roles by k-means on average position; 68–74% agreement with scouts	The principle that metric weights should be learned from outcomes, not asserted — the target for P1, not achievable from season aggregates. Also: it does not handle correlated features, which v2 does
football2vec (Magdaci)	Word2Vec over event sequences → player embeddings → cosine similarity	Nothing usable: needs full event data (StatsBomb open only)
Public PCA→k-means scouting repos	z-score season stats → PCA → k-means or KNN, cosine or Euclidean	The failure mode to test: snapshot, no possession adjustment, redundancy uncontrolled, clusters asserted. §6 below tests whether clusters exist at all
Transfer Flow corpus (37 episodes, 349k words)	Not a method — but a record of which numbers a working practitioner reaches for	Shots (303 mentions), xG (196), pressing (190), dribbles/carries (79), minutes and availability (457), age (144); always position-relative percentiles — "97th percentile xG for his position, 96th shots, 98th dribble-and-carry OBV." The Salah-decline example uses exactly shots/90, box touches/90 and passes into the box — the template

3. What v2 does

Metrics = the StatsBomb striker + wide-attacker template, mapped to FBref. Two honest substitutions: FBref stopped publishing pressures in 2022, so att-3rd tackles + interceptions stands in for pressures and pressure regains; turnovers = miscontrols + dispossessed. 12 metrics: npxg, shots, npxg_per_shot, box_touches, shot_touch_pct, xa, aerials_won, dribbles_succ, turnovers, press_proxy, fouls_won, box_crosses.
Per 90, then z-scored across all 491 big-5 attackers with 900+ minutes in 2024/25.
PCA-whitened before measuring distance. Keep the components explaining 95% of variance (8 of 12), divide each by its standard deviation. Distance in this space is Mahalanobis distance: two metrics that say the same thing count once, not twice. This is the fix for v1's double-counting.
Similarity = Euclidean distance in the whitened space. The 2-D picture uses the first two raw components and is for reading only.
k-means on the same space, k chosen by silhouette — to test whether archetypes exist, not to assert them.

4. Redundancy — the correlation matrix

	npxg	shots	npxg_per_shot	box_touches	shot_touch_pct	xa	aerials_won	dribbles_succ	turnovers	press_proxy	fouls_won	box_crosses
npxg	1.00	0.76	0.58	0.37	0.04	-0.03	0.19	-0.09	0.08	-0.22	-0.04	-0.39
shots	0.76	1.00	0.17	0.42	0.07	0.18	0.10	0.07	0.13	-0.12	0.07	-0.21
npxg_per_shot	0.58	0.17	1.00	0.13	-0.01	-0.21	0.20	-0.16	0.01	-0.17	0.02	-0.34
box_touches	0.37	0.42	0.13	1.00	-0.55	0.17	0.28	0.47	0.62	0.22	0.36	-0.07
shot_touch_pct	0.04	0.07	-0.01	-0.55	1.00	-0.09	-0.22	-0.33	-0.46	-0.28	-0.36	-0.05
xa	-0.03	0.18	-0.21	0.17	-0.09	1.00	-0.35	0.34	-0.04	0.16	0.04	0.53
aerials_won	0.19	0.10	0.20	0.28	-0.22	-0.35	1.00	-0.08	0.33	0.01	0.20	-0.28
dribbles_succ	-0.09	0.07	-0.16	0.47	-0.33	0.34	-0.08	1.00	0.55	0.33	0.38	0.25
turnovers	0.08	0.13	0.01	0.62	-0.46	-0.04	0.33	0.55	1.00	0.31	0.58	-0.06
press_proxy	-0.22	-0.12	-0.17	0.22	-0.28	0.16	0.01	0.33	0.31	1.00	0.29	0.21
fouls_won	-0.04	0.07	0.02	0.36	-0.36	0.04	0.20	0.38	0.58	0.29	1.00	0.03
box_crosses	-0.39	-0.21	-0.34	-0.07	-0.05	0.53	-0.28	0.25	-0.06	0.21	0.03	1.00

No pair exceeds |r| = 0.8 in the template, which is one reason the template is better than v1's 28. Whitening handles what correlation remains.

5. Sensitivity — how much the answer depends on the metric list

Same player, top-8 neighbours under v1 (28 metrics, raw z) versus v2 (12-metric template, whitened):

Vlahović2 of 8 neighbours in common. Castellanos survives; the rest change.Kolo Muani5 of 8 in common — Juanmi, Akor Adams, Cardona persist.Jonathan David6 of 8 in common.Openda4 of 8 in common.

Read this as: the specific names are metric-dependent; the neighbourhoods are not. Vlahović's list is penalty-box strikers under both definitions (v1: Roque, Dovbyk, Castellanos, Watkins, Haaland; v2: Mosquera, Castellanos, Piccoli, Jiménez, Cutrone). Kolo Muani's is hybrids and wide forwards under both. Any single "his closest comp is X" claim from this kind of model deserves the same scepticism you brought to v1.

Vlahović ↔ Kolo Muani distance: 74th percentile of all attacker pairs in v1, 92th percentile in v2. The headline gets stronger under the template.

6. Do archetypes exist? Testing the k-means story

Silhouette by k = 2…8: 0.24, 0.13, 0.12, 0.14, 0.15, 0.15, 0.15. Best is k = 2 at 0.24 — weak structure. The two "clusters" are 444 players and a rump of 47 with very few box touches and a high shot-to-touch ratio: long-range shooters and wide players mislabelled FW, not an archetype.

There are no natural clusters among big-5 attackers in this space. Quadrant labels — "mace, sword, spear, shield" — are a reading aid drawn over a continuum, not something the data produces. Vlahović, Kolo Muani and David all land in the same cluster. This is why the engine is nearest-neighbour distance, and why cluster labels, if ever shown, are captions rather than findings.

7. Does similarity mean anything? A test against next season

For the 276 attackers with 900+ minutes in both 2023/24 and 2024/25: place each player in the 2023/24 space, take his 8 nearest neighbours, and predict his 2024/25 npxG/90 as the mean of what those neighbours did in 2024/25.

Predictor of 2024/25 npxG/90	RMSE	corr
Positional mean (no information)	0.157	—
Player's own 2023/24 npxG/90 (persistence)	0.124	0.68
Mean of 8 nearest neighbours' 2024/25	0.146	0.42
50/50 blend of the two	0.121	—

Similarity carries real information about future output — and much less than a player's own history. Neighbours beat knowing nothing (0.146 vs 0.157) and lose clearly to persistence (0.124). Blending helps slightly. So: the map is a candidate generator — who plays like him — not a projection. Projection stays with Marcel and the P1 model.

8. v2 neighbour lists

Nearest to Vlahović

Player	Club 24/25	Dist	npxG/90	Shots	Box touches	Aerials	Dribbles	xA	Press proxy	Turnovers
Dušan Vlahović (self)	Juventus	0.00	0.46	3.7	1.8	0.5	0.1	0.05	0.1	1.0
Daniel Mosquera	Hellas Verona	1.42	0.43	3.7	0.9	0.7	0.3	0.06	0.0	1.0
Valentín Castellanos	Lazio	1.45	0.43	4.0	1.5	0.8	0.2	0.10	0.2	0.8
Roberto Piccoli	Cagliari	1.50	0.35	2.8	0.8	0.4	0.1	0.01	0.1	0.8
Raúl Jiménez	Fulham	1.51	0.35	3.4	0.9	0.2	0.0	0.08	0.1	0.6
Patrick Cutrone	Como	1.53	0.28	3.0	1.5	0.3	0.1	0.09	0.2	1.1
Mostafa Mohamed	Nantes	1.59	0.40	3.1	0.6	0.4	0.0	0.03	0.1	0.4
Nikola Krstović	Lecce	1.68	0.29	3.9	0.9	0.4	0.2	0.05	0.1	0.6
Georginio Rutter	Brighton	1.70	0.25	2.7	1.1	0.4	0.3	0.09	0.1	0.9

Nearest to Kolo Muani

Player	Club 24/25	Dist	npxG/90	Shots	Box touches	Aerials	Dribbles	xA	Press proxy	Turnovers
Randal Kolo Muani (self)	Juventus/Paris S-G	0.00	0.46	2.4	1.2	0.4	0.3	0.08	0.1	1.1
Juanmi	Betis/Getafe	1.72	0.34	1.8	0.8	0.4	0.4	0.03	0.2	0.7
Emanuel Emegha	Strasbourg	2.68	0.67	2.4	0.9	0.4	0.2	0.03	0.1	0.9
Irvin Cardona	Espanyol/Saint-Étien	2.80	0.26	1.8	0.6	0.2	0.1	0.15	0.2	0.6
Akor Adams	Montpellier/Sevilla	2.81	0.35	2.1	1.8	1.0	0.1	0.01	0.0	1.0
Arnaut Danjuma	Girona/Villarreal	2.84	0.26	2.1	1.5	0.1	0.3	0.13	0.2	0.8
Mario Götze	Eint Frankfurt	2.89	0.17	0.5	0.5	0.0	0.0	0.11	0.1	0.3
Orri Steinn Óskarsson	Real Sociedad	2.94	0.50	1.8	0.8	0.5	0.0	0.05	0.1	0.6
Myron Boadu	Bochum	3.04	0.60	2.3	1.0	0.1	0.1	0.14	0.0	0.9

Nearest to Jonathan David

Player	Club 24/25	Dist	npxG/90	Shots	Box touches	Aerials	Dribbles	xA	Press proxy	Turnovers
Jonathan David (self)	Lille	0.00	0.38	2.3	1.2	0.1	0.2	0.11	0.1	0.6
Takumi Minamino	Monaco	0.93	0.36	2.3	0.8	0.2	0.1	0.19	0.1	0.7
Danny Welbeck	Brighton	0.97	0.37	2.5	1.2	0.3	0.2	0.06	0.1	0.8
Gianluca Gaetano	Cagliari	1.01	0.44	2.7	1.1	0.1	0.3	0.21	0.1	0.7
Nicolas Jackson	Chelsea	1.04	0.50	3.1	1.4	0.1	0.2	0.08	0.1	1.1
Fábio Silva	Las Palmas	1.07	0.31	2.2	1.2	0.2	0.2	0.09	0.1	0.9
Arnaud Kalimuendo	Rennes	1.09	0.31	2.4	0.8	0.2	0.1	0.07	0.1	0.7
Joshua King	Toulouse	1.25	0.36	2.7	0.7	0.2	0.1	0.09	0.1	0.4
Alexander Isak	Newcastle Utd	1.26	0.56	3.1	0.7	0.1	0.1	0.12	0.1	0.5

Kolo Muani's nearest neighbour sits at 1.72 and the next at 2.68 — his second-closest is further away than Vlahović's eighth. He is a genuinely unusual profile in this space, not a mid-table striker.

6b. Correction: profiles exist as regions, not as gaps

§6 asked whether there are gaps between groups (silhouette). The better question is whether there are stable, nameable regions — which a continuum can have. Test: fit k-means on 2022/23 + 2023/24 pooled (both seasons complete), assign each season separately, and ask whether players stay in their cluster.

k	Same cluster, year to year	Chance	Adjusted Rand	Bootstrap stability
2	80%	50%	0.36	0.97
3	75%	34%	0.38	0.91
4	76%	34%	0.42	0.76
5	62%	28%	0.32	0.57
6–8	≤51%	≤23%	≤0.23	—

Three persistent attacking profiles. Players stay in their profile at more than double chance, and the structure collapses beyond k=4 (one of the four is an empty degenerate cluster caused by the shot-touch-% tail). So: no gaps (§6 stands), but three real regions (§6 was wrong to imply none). Any claim of seven or more archetypes is naming noise.

Profile	Signature (z)	2023/24 examples
Box 9 (n=173)	npxG +0.8, npxG/shot +0.6, aerials +0.6; crosses −0.6, pressing −0.6, dribbles −0.6	Solanke, Watkins, Osimhen, Zapata, Lautaro, Haaland, Vlahović, Kolo Muani, David, Openda
Wide dribbler (n=144)	dribbles +1.1, fouls won +0.7, turnovers +0.6; aerials −0.4	Saka, Foden, Kvaratskhelia, Sávio, Gordon, Leão, Lookman, Soulé
Second-line creator (n=180)	pressing +0.6, box crosses +0.6, xA +0.1; box touches −0.8, npxG −0.8, shots −0.8	Gibbs-White, Bruno Fernandes, McGinn, McNeil, Candreva

The nuanced Kolo Muani answer: same family as Vlahović, opposite corner of it. On clean 2023/24 data their distance is at the 100th percentile of all pairs; Vlahović's nearest neighbour is Osimhen at 0.86, Kolo Muani's is Maupay at 3.06. The family label alone would have hidden the second fact.

6c. Finer grain: a persistence-selected taxonomy over seven seasons

Seven complete seasons (2017/18–2023/24; 3,533 attacker-seasons, 1,478 players) give six year-pairs to test persistence on. Method: 25 per-90 metrics (the template plus zone touches, progression, receiving, passing types), audited for season consistency — xA and progressive passes were dropped because FBref only populates them from 2023/24, which had been silently pulling every 2023/24 season into its own cluster — then standardised within season, whitened, and clustered hierarchically: families first, then sub-profiles inside each family, with the number of groups at each level chosen by year-to-year stability, not by silhouette.

Level	k	Adjusted Rand, year to year	Same group	Chance
Families	3	0.51	79%	44%
Sub-profiles within the wide family	2	0.38	81%	52%
Sub-profiles within the 9 family	2	0.39	81%	52%
Any third split within a family	3+	fails the floor (ARI ≥ 0.25 and stability ≥ 1.5× chance)

Four persistent attacking profiles, plus a transient bin. The wide/9 boundary is very firm — almost nobody crosses it year to year. Each family splits once, and only once, into two stable sub-profiles. Anything finer does not survive a season.

Profile	Share	Stickiness	Signature (per 90)	Who lives here, 2021/22–2023/24
Finishing 9	22%	69%	npxG 0.39, box touches 4.8, shots high; passing and carrying volume low	Kane, Watkins, David, Salah, Lewandowski, Lautaro, Haaland, Osimhen, Immobile, Ben Yedder
Target 9	16%	66%	aerials won 3.8 (vs ~1 elsewhere), npxG 0.30, turnovers high, carrying lowest	Toney, Solanke, Joselu, Ünal, Muriqi, Pinamonti, Budimir, Nzola, Mounié
Wide creator	19%	70%	crosses 5.3, key passes high, final-third passing; npxG 0.18	Bourigeaud, Grifo, Foden, Brandt, Bruno Fernandes, Bowen, Berardi, Son, Candreva, Dybala
Wide carrier	35%	69%	progressive carries and carry distance high, take-ons 3.5, pressing; npxG 0.18	Leão, Vinícius, Rodrygo, Saka, Sterling, Mbappé, Bernardo Silva, Ocampos, Almirón
Atypical (not a profile)	7%	10%	Rate-stat extremes on low volume — very high pass completion on few passes, long shot distance, high aerial % on few duels	Where a player lands in an unusual season: Vlahović 21/22, Kolo Muani 23/24, Palmer 23/24, Weghorst, Zapata

Movement between profiles is informative in itself. Of 386 finishing-9 seasons with a next season, 61 became target-9 and 34 became wide carriers; of 309 target-9 seasons, 61 became finishing 9s. Fewer than 5% of wide seasons became any kind of 9. The permeable boundaries are finishing↔target and creator↔carrier; the wide/9 boundary is close to a wall.

Profile trajectories, Juventus cases

Player	Season by season	Reading
Vlahović	Target 9 (Fiorentina 19/20, 20/21) → atypical (21/22, the split season) → Finishing 9 (Juve 22/23, 23/24)	Became less of a target man at Juve; ended as a conventional finishing 9
Kolo Muani	Finishing 9 (Nantes 20/21, 21/22) → Target 9 (Frankfurt 22/23) → atypical (23/24 Frankfurt/PSG)	A different profile every season; the only headline case with no stable identity — consistent with every earlier finding
Jonathan David	Finishing 9 all four seasons	The incumbent is the stable version of what Kolo Muani was at Nantes
Openda	Finishing 9 (Lens, Leipzig)	Same profile as David — Juve bought two of the same thing in 2025
Chiesa	Wide carrier (Fiorentina) → finishing 9 (19/20) → wide creator (Juve 23/24)	A genuine role change across clubs
Salah · Haaland · Lautaro · Osimhen	Finishing 9 every season on record	Elite output is stable in profile as well as in level
Leão	Finishing 9 (Lille, early Milan) → wide carrier (21/22 onward)	Milan moved him wide and he stayed there

What this taxonomy is for. Candidate generation and a sanity check on fit — "Juve replaced a finishing 9 with a player who has not held any profile for two consecutive seasons" is a statement the data supports. It is not a quality measure: the finishing-9 profile contains Haaland and Gyasi. Quality is Marcel and the P1 model; profile is where to look.

8b. Neighbours on clean data — 2023/24

Nearest to Vlahović (Juventus, 2,307 min)

Player	Club	Dist	npxG/90	Shots	Box touches	Aerials	Dribbles	xA
Victor Osimhen	Napoli	0.86	0.52	3.9	6.9	1.7	0.7	0.04
Isaac Bernal	Sevilla	1.27	0.63	3.6	5.5	1.4	1.1	0.09
Luis Javier Suárez	Almería	1.30	0.56	3.8	5.5	1.0	1.2	0.05
Folarin Balogun	Monaco	1.61	0.45	3.4	7.4	0.4	0.6	0.10
Akor Adams	Montpellier	1.71	0.50	3.2	4.7	2.2	0.6	0.06
Takuma Asano	Bochum	1.72	0.36	2.9	4.3	0.7	0.7	0.04
Wahbi Khazri	Montpellier	1.88	0.26	3.3	4.1	0.7	1.1	0.06
Maximilian Beier	Hoffenheim	1.89	0.42	3.0	5.4	0.7	1.2	0.11

Nearest to Kolo Muani (Frankfurt/PSG, 1,421 min)

Player	Club	Dist	npxG/90	Shots	Box touches	Aerials	Dribbles	xA
Neal Maupay	Brentford/Everton	3.06	0.45	2.6	5.0	0.9	0.6	0.14
Samuel Omorodion	Alavés/Granada	4.03	0.52	3.0	5.0	2.7	1.0	0.03
Facundo Pellistri	Granada/Man Utd	4.08	0.21	1.3	4.0	1.0	1.1	0.10
David Datro Fofana	Burnley/Union Berlin	4.48	0.42	2.6	5.0	1.3	2.0	0.03
Portu	Getafe/Girona	4.89	0.25	1.5	4.1	0.6	0.3	0.18
Brennan Johnson	Forest/Spurs	4.95	0.41	2.1	6.3	0.2	1.0	0.17

Note the distances: Vlahović's eighth neighbour is closer than Kolo Muani's first. That is the outlier signature, and it survives the correction.

9. What is still not done

Possession adjustment. StatsBomb adjusts pressures, tackles and interceptions for team possession. FBref's player tables don't carry team possession; it needs a join to team-level data. Until then the pressing proxy favours players on low-possession teams.
Team-context residualisation — the spec's §6.2 requirement. A striker at a possession-dominant club drifts toward the creator side of the map for reasons that are not his.
2024/25 and 2025/26. The archived release truncates 2024/25 for half the template (possession/defence/misc ≈ 20% of the season) and has 3–5 games of 2025/26. Both seasons need a live FBref scrape before this map can be current.
Learned weights. PlayeRank's lesson — learn what matters from outcomes — is the P1 model's job. This map weights every template metric equally after whitening.
Beyond the big five. Opta-grade metrics stop at the big five and a short list; a feeder-league map has to be built from Sofascore, separately, never mixed.

player_recruitment · method v2 · data: FBref 2024/25 via worldfootballR (archived release) · code: scratchpad/fb/sim2.R · results: experiments/style-space-v2-template.json