But video is expensive. Time-expensive. Attention-expensive. You can’t scout every team, every match, every season this way.

That’s where data earns its place — not to replace the eye test, but to make decisions efficient. And if someone tells you “data is useless, only the match matters,” what they really mean is they haven’t found the right data to reflect what the match is showing.

Recently, I tested this idea on Arsenal’s 2024–25 season.

Using only public match data from Sofascore — no Opta, no StatsBomb, just freely available stats — I broke down Arteta’s open-play structure into three distinct attacking patterns:

• Pattern A · Right-Side Overload (~55%, primary) — White inverts high, Ødegaard occupies the right halfspace, Saka isolates the left-back. The chain ends in a cross or cut-back to the centre forward.
• Pattern B · Left Isolation / Underlap (~25%) — Martinelli pinned wide, Calafiori underlaps, low cut-back to Merino.
• Pattern C · Vertical Transition (~20%) — Rice and Raya spring play vertically; Havertz drops to link, the wingers attack the channels.

For each pattern I drew three views: a role-link map (the shape of the pattern), a per-player focus grid (each role’s linkages in isolation), and a role-execution grading pitch (how well each role actually performed its task, with the chain’s bottleneck ringed in black).

Pattern A · Right-Side Overload (~55%)

The primary pattern. The heavy links live on the right: Saliba’s carry into White and Ødegaard, the Saka–Ødegaard–White triangle, then a cross or cut-back. The per-player grid shows how concentrated the structure is — Ødegaard and Saka’s panels are dense, Martinelli’s nearly empty.

Grading the execution role by role: Saliba A+, Rice A+, Saka A+, Ødegaard A. The build and progression phases are elite. Then the chain reaches the box.

The bottleneck — the dark ring — sits on the centre forward. Havertz grades C: big-chance conversion poor. Every upstream role does its job; the last one doesn’t.

Pattern B · Left Isolation / Underlap (~25%)

The mirror-image idea: Martinelli held wide for the 1v1, Calafiori underlapping into the channel, a low cut-back arriving for Merino. The link weights show the Martinelli–Calafiori axis carrying the whole pattern.

This pattern has two breakers, and neither is in the build-up.

Martinelli grades C (1v1 success and finishing down across 24–26), and Calafiori grades C (underlap dynamism limited). The back seven all grade A or better — the platform is fine; the left-side blade is blunt. Havertz, on the near-post pin, again grades C.

Pattern C · Vertical Transition (~20%)

The direct route: Raya or Rice plays the first vertical pass, Havertz drops to link, Saka and Martinelli run the channels, the 8s arrive late.

Raya B+, Rice A+ on the first pass, Saka A on depth runs. The hinge of the whole pattern is the dropping striker — and that is exactly where it breaks.

Havertz, the designated link, grades C: link play and spin inconsistent. The launch pad is elite, the runners are ready, the connection point fumbles.

The Finding

For each pattern, I mapped all 11 roles’ tasks and graded execution. The finding was striking: in two of three patterns, the chain broke at the same role — centre forward. The structure built elite chances, but the finisher didn’t convert. In the third, the same player broke the chain in a different function.

The data-driven conclusion was simple: Arsenal’s number-one recruitment priority had to be an elite, ruthless No. 9.

This summer, Arsenal signed Viktor Gyökeres.

And this season, they ended a 22-year wait to lift the Premier League title.

The data didn’t replace the football. It pointed exactly where the football was already telling us to look — just faster, cheaper, and more communicable.

If you find yourself saying “the data doesn’t capture what I’m seeing,” the answer isn’t to abandon the data. It’s to design better data.

Data: Sofascore public match data · Charts: synthetic role-link weights, style adapted from The Athletic’s pass-network charts · Stack: Python + matplotlib

I used StatsBomb’s open event data for the 2022 World Cup Final (Argentina 3–3 France) to build eight passing visualisations, trying to answer one question:

How much of a team’s “style of play” can you actually recover from the passing data of a single match?

Why Passing Data, and Why Is It So Hard to Draw?

Nobody disagrees that passing data matters. A pass is the most frequent event in football — this final alone produced 1,263 of them — and every pass carries a start point, an end point, a height, a length, a direction, and an outcome. That richness is exactly the problem. The moment you draw a chart, the questions multiply: length, height, or direction? Start-point density or the spatial distribution of average length? What goes on the colour axis — frequency, length, forwardness? What does opacity mean?

There is no canonical answer. Every figure is one projection of the same event stream, and a different projection tells a different story. A heatmap of pass origins says “where did they have the ball”; the same cells coloured by mean pass length says “what did they do with it from there”; colour them by forwardness and you get intent. Same data, three different matches on the page.

The longer I do this, the more I believe a visualisation’s real value is not to replace an expert watching the tape, but to take the judgment they build after hours of re-watching and surface it as quickly and intuitively as possible — giving subjective expertise an objective substrate that can be shared, debated, and falsified.

A good chart isn’t an answer. It’s evidence.

The Figures

All plots use StatsBomb open event data, rendered with Python, matplotlib and mplsoccer. Both teams attack left → right; each pitch shows passes by one team, smoothed with a Gaussian kernel (σ ≈ 1.8 m) so the colour reads as a local average rather than individual dots.

Pass height. Colour is the average StatsBomb height category per cell (ground → low → high), opacity is local pass density.

Pass direction. Colour is the average forwardness of passes starting in each cell: +1 fully forward, 0 sideways, −1 backward.

Pass length. Colour is mean pass length per cell, opacity again is density — so a bright yellow blob means “from here, they hit it long, often.”

Distributions. Overlaid pass-length histograms and CDFs, plus each team’s length distribution split by direction sector (forward / sideways / backward).

So what did these figures say about this final?

1. Argentina Played the “Possession” Version of a Final

693 passes vs France’s 570. Mean length 20.2 m (France 22.2). 69% on the ground. Pass-start density bright across midfield and the attacking third (164 attacking-third passes). On the forwardness field, “forward” hotspots are scattered everywhere — patient build-up that can launch from anywhere.

That last point is the one I would not have guessed from memory. The popular recollection of this final is Messi-centric: moments, not structure. The data says the structure was there all along — Argentina sustained possession deep into France’s half, recycled it short, and progressed from many different zones rather than through one designated outlet. The shape of the histograms backs this up: Argentina’s distribution piles up at 10–20 m, and their backward passes are the shortest in the match (mean 15.7 m) — quick resets to keep the machine turning, not panicked clearances.

2. France Played the “Transition” Version of a Final

41% of France’s defensive-third passes were ≥ 25 m long balls (Argentina: 31%). High-pass share 21% vs 18%, clustered on the left of their own half — Upamecano and Varane spraying forward. The mean length of France’s forward-direction passes was 25.7 m, three metres longer than Argentina’s. And only 100 attacking-third passes — almost the entire attack relied on a handful of direct, vertical transitions.

You can see it in one glance at the pass-length heatmap: the brightest yellow on France’s pitch sits squarely around their own penalty area and left channel. The launch zone. Argentina’s equivalent area is dimmer and more purple — shorter, calmer exits. France’s by-direction histogram makes the same point a different way: their forward passes have a visible second bump out at 55–70 m that Argentina’s simply doesn’t have. Those are the balls hunting Mbappé’s shoulder run.

3. The Forwardness Field Tells the Subtlest Story

Both teams have nearly identical mean forwardness (+0.22 vs +0.24). If you only read the summary statistic, you’d call them equally “direct” and move on. But the spatial distribution is opposite: Argentina’s forward-red is scattered all over the pitch; France’s is compressed into the defensive flanks plus right midfield. Tchouaméni and the centre-backs were the launch pad.

This is exactly the kind of finding that justifies drawing a map instead of quoting a mean. Two identical averages, two completely different footballing ideas underneath. It’s also a small warning about single-number team metrics in general: directness, verticality, tempo — every one of them is a spatial pattern flattened into a scalar, and the flattening is where the style gets lost.

What the 3–3 Actually Means

Argentina tried to grind it out through possession. France tried to steal it through Mbappé’s speed in transition. One patient, one vertical. 3–3 plus penalties is the cleanest evidence that neither philosophy could suppress the other.

The usual caveats apply, and they matter. This is one match — the highest-variance, highest-stakes match there is — and a final is a terrible sample of a team’s “true” style. France spent long stretches chasing the game, which mechanically inflates directness; Argentina spent long stretches protecting a lead, which inflates possession share. Event data also can’t see the off-ball: the press that forced a long ball doesn’t appear in the passing table, only its consequence does. A one-game read is a hypothesis, not a verdict.

But within those limits, the answer to my opening question is: more than I expected. Eight projections of a single match’s passing recovered, without any tactical annotation, the same characterisation a scout would give you — Argentina’s distributed patience, France’s compressed verticality, the centre-back launch pad, the Mbappé outlet. The data didn’t discover anything the experts didn’t know. It made what they know visible, fast, and checkable — which is the whole job.

A good visualisation isn’t there to make analysts look sophisticated. It’s there to turn a pattern only a few experts can see on tape into something almost anyone can grasp in 30 seconds. Data is cold; good figures amplify the warm, intuitive things experts already feel.

Data: StatsBomb open-data · Stack: Python + matplotlib + mplsoccer

The Idea

A goal is a rare, noisy event. A team can play well and lose; a single deflected shot can decide a match. xG smooths out that noise by scoring each chance on its quality rather than its outcome. Sum a team’s xG over a match and you get a better estimate of how many goals they should have scored than the scoreline alone.

The Features That Matter

Most of the predictive power in a basic model comes from a handful of features:

• Distance to goal — the strongest single predictor.
• Angle to goal — how much of the goal mouth the shooter can actually see.
• Body part — header vs. foot.
• Play pattern — open play, set piece, fast break, penalty.

You can compute distance and angle directly from the shot’s (x, y) coordinates in event data. StatsBomb’s open data is a good place to start.

A Minimal Model

import pandas as pd
from sklearn.linear_model import LogisticRegression

# shots: DataFrame with x, y, is_header, is_open_play, goal (0/1)
shots["distance"] = ((120 - shots.x) ** 2 + (40 - shots.y) ** 2) ** 0.5

X = shots[["distance", "is_header", "is_open_play"]]
y = shots["goal"]

model = LogisticRegression(max_iter=1000).fit(X, y)
shots["xg"] = model.predict_proba(X)[:, 1]

Logistic regression is the right starting point: the output is already a probability, the coefficients are interpretable, and it’s hard to overfit. Once this baseline works, gradient-boosted trees usually add a few points of calibration — but get the baseline honest first.

Reading the Output

The number to trust is not any single shot’s xG — it’s the aggregate. Over a season, a striker who consistently outperforms their xG is either an elite finisher or due for regression. Telling those two apart is where the analysis starts, and it’s a theme I’ll keep coming back to.

This is a starter post for the Data & Models track — more on calibration, tracking data, and possession value to come.