Understanding Football Through Data

Why Football Analytics Matters

Football is a low-scoring, fluid, invasion game played by twenty-two people making thousands of decisions over ninety chaotic minutes. That combination makes it both the hardest team sport to quantify and the one where good measurement pays the largest dividend.

For most of the sport's history, the scoreboard was the only number that mattered, and it lied constantly. A team could dominate, hit the post three times, concede a deflected goal and "lose 1–0" — a result the raw scoreline records as a deserved defeat. Analytics exists to close the gap between what happened and what was likely to happen, so that clubs make decisions on the underlying process rather than on noisy, small-sample outcomes.

The reason this matters so much in football specifically is the goal. Goals are rare — roughly 2.7 per match across Europe's top leagues — which means a single fortunate or unfortunate event can swing a result, and a season is short enough (typically 34–38 league games) that luck does not fully wash out. A striker who scores 15 from 12 expected goals looks elite; the data tells you to expect regression. A pressing system that concedes few shots but loses on the day is, over time, very likely sound. Analytics is the discipline of separating signal (repeatable skill and tactics) from noise (variance you cannot bank on).

From Moneyball to the modern recruitment department

The cultural reference point is baseball's Moneyball, but football's analytics revolution arrived later and looks different. Baseball is a sequence of discrete one-on-one events that decompose cleanly into individual statistics. Football is continuous and deeply interdependent: a full-back's overlap only "works" because of where the winger and the opposition's defensive line are standing. The field had to invent new tools — expected goals, possession-value models, tracking-data physics — precisely because borrowed baseball thinking did not transfer. Today, every elite club runs a data department, recruitment is routinely informed by models, and opposition analysis is a blend of video and numbers.

Who this handbook is for

How to read this handbook

The four parts build on each other. Part 1 establishes what data is and how it is collected. Part 2 covers the advanced metrics and models that turn data into meaning. Part 3 is the hands-on toolkit — workflow, Python, visualization and video. Part 4 ties it together with realistic, worked case studies, ending in a full tactical report. You can read linearly or jump via the sidebar; cross-references are linked throughout.

1.1 The Role of a Football Data Analyst

"Football data analyst" is not one job. It is a family of roles that share a skill set but point at very different decisions — from the recruitment meeting to the training-ground whiteboard to the betting trading desk.

At its core the role is translation. On one side sits messy, high-volume data; on the other sit football people — coaches, sporting directors, scouts — who need a clear answer to a concrete question. The analyst turns "we have 38 matches of event data" into "their right-back steps out to press and leaves space in behind; we should target it with early diagonals." The value is never the chart. It is the decision the chart changes.

The main flavours of the role

What the job actually demands

Three competencies sit underneath all of these, and the best analysts are strong in each: football understanding (you must know why a stat means something tactically), technical skill (querying, cleaning and modelling data — typically in Python, SQL and a BI/visualization tool), and communication (a brilliant model that nobody acts on is worthless). A common beginner mistake is to over-index on the technical axis. In a club, a clear one-slide answer beats an elegant notebook every time.

1.2 Data Literacy and a Scientific Approach to Data

Data literacy is not about memorising metrics. It is the habit of asking where a number came from, what it can and cannot say, and how confident you should be — the scientific method applied to a football pitch.

The analytical loop

1. Question

   Start with a football question, not a dataset. "Are we creating enough high-quality chances from open play?" is researchable. "Let's look at the data" is not.

2. Hypothesis

   State what you expect and why. "I think our low chance volume is a build-up problem, not a finishing problem." A hypothesis you can be wrong about is the engine of good analysis.

3. Evidence

   Gather the right data at the right grain. Match the metric to the question and be honest about sample size.

4. Test

   Compare against a baseline — league average, the opponent, your own past form. A number without a reference point is meaningless.

5. Communicate

   Deliver the answer in the language of the decision-maker, with the uncertainty attached.

Three habits that separate professionals from dashboards

Always ask about sample size. Football's low event-rate means most single-match numbers are noise. One game of xG tells you almost nothing; ten games starts to. Per-90 rates from 200 minutes of play are dangerous. The professional instinct is to distrust extreme values from small samples and to favour rolling windows.

Correlation, causation and confounders. Teams that press high tend to win more — but elite teams both press high and have better players, so the press is partly a marker of quality, not only a cause of it. A scientific analyst names the confounder before drawing a conclusion.

Distributions over averages. "Average shot distance: 18m" hides whether a team takes many close-range chances and a few hopeful long-rangers, or a steady stream of 18m efforts. The shape of the data usually carries the tactical story; the mean often erases it.

1.3 Data Types and Data Collection Systems

Everything downstream — every metric, model and report — is constrained by the data underneath it. Knowing the four major data types, and their blind spots, is the single most useful piece of literacy in the field.

The four layers of football data

Event data, in detail

Event data is what most analysts spend their day in. A provider's operators (or, increasingly, machine-vision systems with human verification) tag every on-the-ball action. A single shot event might carry: the player, the team, the minute and second, the x/y location, the body part, the play pattern (open play, corner, fast break), whether it was a "big chance," and — crucially for modelling — the position of every other player at the moment of the shot (a freeze-frame) in the richest feeds.

Its limitation is right there in the name: event data only records moments when the ball is touched. It is blind to the run a striker made that dragged a centre-back away, or the cover-shadow a midfielder cast to block a passing lane. For that, you need tracking.

Tracking data and the off-ball game

Tracking data turns football into geometry. With 22 player coordinates ten to twenty-five times a second you can compute distances between lines, the compactness of a block, how quickly a press collapses space, who is free, and which zones a team controls. It powers the most advanced models in the sport — pitch control, off-ball value, defensive line-height analysis. Two collection methods dominate:

• Optical / in-stadium — fixed multi-camera rigs track every player. Highest accuracy, but requires installation and is club-specific.
• Broadcast computer vision — companies such as SkillCorner extract positional and physical data straight from the TV feed, with no sensors required. This democratised tracking: you can now get running and speed numbers for opponents you will never install hardware for. The trade-off is occasional gaps (players off-camera) and slightly lower precision.

The provider landscape (2025–26)

The market consolidated sharply. Hudl acquired Wyscout, then InStat, and most recently StatsBomb, bundling video scouting, event data and player-location data. Opta (Stats Perform) remains the data that powers much of the industry and broadcast. SkillCorner leads broadcast-derived tracking and physical data. IMPECT, known for its packing data, was acquired by Catapult in late 2025. For learning and portfolios, FBref (free, Opta/Stats Perform-powered) and StatsBomb's free Open Data are the standard entry points.

1.4 A Day in the Life of a Professional Football Data Analyst

The job is far less glamorous and far more rhythmic than outsiders imagine. It runs on the fixture calendar, and most of it is preparation for a forty-five-minute conversation that has to land.

Below is a representative match-week day for a club opposition analyst two days before a fixture (often called "MD-2," matchday minus two).

On matchday the rhythm changes: live tagging or live data monitoring, half-time numbers to inform the coach's talk, and a rapid post-match summary. The day after is for the post-match deep-dive (see 4.3). The throughline is that data work in a club is never an end in itself — it is always scaffolding for a coaching decision, delivered on a deadline set by the next kick-off.

"The analyst's job is to make the coach's next decision a little less of a guess — and to do it before the bus leaves."A working principle of club analysis departments

2.1 Introduction to Advanced Metrics

Traditional stats count what happened. Advanced metrics estimate what should have happened, and how much each action moved a team toward a goal. That shift — from counting to valuing — is the whole revolution.

A "shot" is a count. It treats a tap-in and a 35-yard hopeful effort as equal. xG replaces the count with a probability: this chance, given its location, angle, body part and pressure, would be scored a certain fraction of the time. Do that for every action — not just shots — and you get a family of models that value passing, carrying, pressing and positioning. Three properties make a good advanced metric:

• Predictive — it should describe future performance better than the raw outcome. Expected goals predict future goals better than past goals do.
• Stable — it should stabilise on a reasonable sample, so it reflects skill, not variance.
• Interpretable — a coach should be able to understand what it rewards. A black-box number nobody trusts gets ignored.

2.2 Shooting Metrics

Shooting metrics answer two separate questions that beginners constantly conflate: how good were the chances? (creation & concession) and how well were they taken? (finishing).

Expected Goals (xG)

xG assigns every shot a probability of becoming a goal, between 0 and 1, based on the characteristics of the chance. A penalty is worth about 0.76 xG; a tap-in might be 0.9; a speculative effort from distance, 0.03. Sum a team's xG across a match and you get a far better estimate of who "deserved" to win than the scoreline. The model is trained on hundreds of thousands of historical shots, learning the conversion rate for each combination of features.

Distance and angle do most of the work — a shot's value collapses quickly as you move away from goal and toward the byline. Richer models add context. StatsBomb's model uses freeze-frames (the position of the keeper and every defender at the moment of the shot) and shot-impact height; Opta's analyses up to ~20 contextual factors per shot but, like most models, has no freeze-frame. This is why the same shot can read 0.12 on one provider and 0.22 on another — there is no single "true" xG, only models with different inputs. Always know whose xG you are quoting.

Reading a shot map

The shot map is the analyst's most recognisable visual. Each shot is a dot, placed where it was taken, sized by its xG and coloured by outcome. At a glance you can read a team's chance quality and profile: a cluster of big dots in the box signals a side that works the ball into high-value areas; a scatter of small dots from distance signals a team that settles for low-probability efforts.

The finishing layer: xG vs goals, and post-shot xG

Compare a player's goals to their xG and you get a finishing signal. Persistently outscoring xG by a wide margin can indicate elite finishing — but over small samples it is usually variance, and the honest default is to expect regression toward the model. To isolate finishing more cleanly, analysts use post-shot xG (PSxG), which is computed only for shots on target and adds where in the goal the ball was heading. PSxG minus xG measures shot-placement skill; for goalkeepers, PSxG minus goals conceded is the standard shot-stopping metric.

2.3 Chance Creation Metrics

Goals are the tip; chance creation is the iceberg. These metrics value the passing, carrying and progression that manufacture shots — and, in their most advanced form, value any action that increases the threat of a possession.

Expected Assists (xA) and beyond

Expected Assists (xA) takes every completed pass that led to a shot and assigns it the xG of the shot it created. It rewards the quality of chance created, not whether the teammate finished — a perfect cut-back that the striker skies still earns its xA. It is to creators what xG is to finishers. Its limitation: it only credits the final pass before a shot. The line-breaking pass two phases earlier, the carry that drew defenders — these get nothing. That gap is what possession-value models fill.

Possession value: xT and VAEP

This is the conceptual heart of modern analysis. Both models answer: how much did this action increase my team's chance of scoring?

Expected Threat (xT)

Introduced by analyst Karun Singh in 2018, xT lays a value surface over the pitch by dividing it into a grid (commonly 16×12 zones). Each zone gets a value: how likely a goal is to result from possession there. Moving the ball from a low-value zone to a high-value one earns the player the difference in zone value. A carry from the centre circle into the box is highly rewarded; a sideways pass between two low-value zones earns almost nothing.

xT is intuitive and easy to communicate, which is why it is the most widely used possession-value model in the industry. Its weaknesses: it values only ball-progression actions (passes and carries), ignores defensive actions and shots, and — because it is usually built on event data alone — it cannot see whether the player was under pressure or whether the pass was risky.

VAEP — Valuing Actions by Estimating Probabilities

VAEP goes further. It values a broad set of actions (including defensive ones) by estimating how each changes two probabilities: the chance your team scores in the next few actions, and the chance your team concedes. An action's value is the gain in scoring probability minus the rise in conceding probability. Crucially, VAEP includes game context — score difference, time remaining, field position — and frames valuation as a machine-learning classification problem. It is implemented in the open-source socceraction library, which first converts any provider's event feed into a common format called SPADL.

2.4 Pressing & Defensive Metrics

Defending is the hardest part of the game to measure, because the best defensive action is often the one that never has to happen. The metrics here are proxies — useful, but to be read with care.

PPDA — Passes Per Defensive Action

PPDA is the most widely used single-number proxy for pressing intensity. It counts how many passes the opponent is allowed to complete before the pressing team makes a defensive action (tackle, interception, challenge or foul), measured in the attacking ~60% of the pitch (the opponent's defensive three-fifths).

A PPDA around 7–8 indicates suffocating, aggressive pressing — the team intervenes after only a handful of opponent passes. A PPDA of 15+ signals a side content to let the opponent have the ball and defend in a mid- or low-block. Neither is "better"; they are tactical choices.

Other defensive lenses

Defensive actions by zone & height of the defensive line. Where a team makes its tackles and interceptions tells you its block height. A high cluster = aggressive front-foot defending; a deep cluster = a low block. From tracking data you can measure the literal height of the defensive line and how it moves.

Packing (Impect). Packing counts how many opponents a pass or dribble takes out of the game by playing past them. Defensively, it measures how many defenders an action bypassed — a way to value the act of not being eliminated. It reframes both passing and defending around the number of players removed from the play.

Expected goals against / prevented (PSxG–GA). The flip side of creation: how many high-quality chances a defence concedes (xGA), and how a keeper performs against them.

Tracking-era defensive value. The frontier — pitch control and off-ball models — credits defenders for the space they deny and the passing lanes they shadow, finally putting a number on the "action that never had to happen."

2.5 Player Functions and Clustering

A "midfielder" can be a destroyer, a metronome, a box-crasher or a deep-lying playmaker. Positional labels are too blunt for recruitment. Clustering lets the data define roles from how players actually play.

From positions to data-driven roles

The traditional position (CB, FB, DM, CM, AM, W, ST) tells you where a player stands, not what they do. Two "centre-backs" might be a ball-playing libero and a no-nonsense stopper — opposite profiles sharing a label. Clustering solves this by representing each player as a vector of per-90, style-based metrics (progressive passes, carries, pressures, aerials, shot volume, defensive actions, xT generated…) and letting an algorithm group players whose statistical fingerprints are similar, regardless of nominal position.

The clustering workflow

1. Select features

   Choose metrics that describe style, not just quality — you want to group a good and an average deep-lying playmaker together, then judge quality separately.

2. Normalise & control for position

   Scale features (e.g. z-scores) so no single metric dominates. Often cluster within position groups, or per-possession-adjust.

3. Reduce dimensions

   Use PCA or similar to compress dozens of correlated metrics into a handful of meaningful style axes and to visualise.

4. Cluster

   Apply k-means (or hierarchical / Gaussian mixtures). Use the elbow method or silhouette score to choose the number of clusters K.

5. Interpret & name

   The hard, football part: look at each cluster's average profile and give it a human label — "ball-progressing 6," "wide creator," "pressing forward."

A caution: clustering groups players by style, not level. A Championship metronome and a Champions League metronome can land in the same cluster. Always pair the role label with a quality and level-of-competition adjustment before drawing a recruitment conclusion. See 4.4 for a worked player analysis, and 3.3 for building radars in Python.

2.6 How Data Shapes Technical and Scouting Decisions

Metrics are only worth the decisions they improve. Here is how the models above feed the three decisions clubs spend most money and emotion on: who to sign, how to play, and whether the process is working.

Recruitment

Modern recruitment is a funnel. Data widens the top of it — instead of watching the handful of players a scout happened to see, a club can filter every player in dozens of leagues against a statistical profile in minutes, then send scouts to watch the shortlist. Possession-value models rank contribution; clustering ensures stylistic fit; age and contract data flag value. The model never makes the final call — it decides who gets watched, which is where most of the leverage is. It also guards against bias: the eye over-weights the last spectacular game; the data remembers the whole season.

Tactics & opposition planning

Data shapes both how a team plays and how it prepares for opponents. A coach can see, with evidence, that the side concedes most of its xG from crosses and adjust the defensive scheme; that the press is bypassed by the opponent's goalkeeper's long kicking and plan a trap; that a winger's xT is highest cutting inside onto his stronger foot. The opponent report in 4.2 is this in action.

Performance review & player development

Post-match, data separates process from result (4.3): did we generate good chances and limit theirs, regardless of the 1–1? For individuals, tracking and event data flag development targets — a full-back whose final-third decision-making lags his physical output, a striker whose movement creates xG he isn't converting.

3.1 The Football Data Analyst Workflow

Behind every clean chart is a pipeline. Professional analysis is repeatable: the same steps, run reliably, every match week. Learn the workflow and the specific tools become interchangeable.

1. Acquire

   Get the data — a provider API, a CSV export, FBref or StatsBomb open data for learning. Know the licence and the schema before anything else.

2. Clean & standardise

   Real data is messy: missing values, inconsistent names, different coordinate systems per provider. Standardising pitch coordinates (e.g. to 0–100) and player IDs is unglamorous and essential.

3. Transform & model

   Aggregate to the grain you need (per-90, per-possession), compute metrics, apply models (xG, xT, VAEP).

4. Visualise & explore

   Pitch plots, shot maps, radars, scatters — both to find the story and to tell it.

5. Communicate

   Package the answer for the decision-maker: a slide, a one-pager, a video with overlaid numbers.

6. Automate

   Turn the one-off into a template. Next week's opponent report should be a parameter change, not a rebuild.

The standard toolkit

3.2 Python Basics for Football

You don't need to be a software engineer. You need to load a table, filter it, group it and compute a metric — 80% of football analysis is exactly that, done well.

The workhorse is pandas, whose core object is the DataFrame — a spreadsheet you control with code. Below, we load a competition's event data with statsbombpy and compute each team's total xG: the single most common starting task in the field.

from statsbombpy import sb
import pandas as pd

# 1 — load events for a single match (StatsBomb free open data)
events = sb.events(match_id=3795506)

# 2 — keep only shots, with the columns we care about
shots = events[events["type"] == "Shot"][
    ["team", "player", "minute", "shot_statsbomb_xg", "shot_outcome", "location"]
].copy()

# 3 — total xG and goals by team  (the core "who deserved it" table)
summary = (
    shots
    .assign(goal=lambda d: (d["shot_outcome"] == "Goal").astype(int))
    .groupby("team")
    .agg(xG=("shot_statsbomb_xg", "sum"),
         goals=("goal", "sum"),
         shots=("shot_statsbomb_xg", "size"))
    .round(2)
)
print(summary)
#            xG  goals  shots
# team
# Arsenal  1.84      2     14
# Chelsea  0.97      1      9

Three steps — filter, derive, group — answer "who created the better chances?" in a dozen lines.

The same three-verb pattern (filter → derive → group) scales to almost everything. Want each player's shots and average shot distance? Group by player instead of team. Want per-90 rates? Join a minutes table and divide. Below, a small helper turns StatsBomb's [x, y] location list into a shot distance — the kind of feature engineering that feeds an xG model.

import numpy as np

GOAL = np.array([120, 40])           # StatsBomb pitch: 120 x 80, goal at x=120

def shot_distance(loc):
    if loc is None: return np.nan
    x, y = loc
    return np.hypot(GOAL[0] - x, GOAL[1] - y)

shots["distance"] = shots["location"].apply(shot_distance)

# average distance of shots that became goals vs. those that didn't
print(shots.groupby(shots["shot_outcome"] == "Goal")["distance"].mean().round(1))
# shot_outcome
# False    18.3      # missed/saved shots taken from further out
# True     11.2      # goals scored from closer in — distance matters

Feature engineering: deriving shot distance from raw coordinates. Notice the result already "rediscovers" why xG weights distance so heavily.

3.3 Data Visualization Basics for Football

In football, the pitch is the axis. A good football visual places data in its spatial context so a coach reads it in seconds. The mplsoccer library makes professional pitch plots straightforward.

Plotting a shot map on a pitch

from mplsoccer import Pitch
import matplotlib.pyplot as plt

pitch = Pitch(pitch_type="statsbomb", line_color="#c4d6c9", pitch_color="#f4f8f5")
fig, ax = pitch.draw(figsize=(8, 5.2))

team_shots = shots[shots["team"] == "Arsenal"]
goals = team_shots[team_shots["shot_outcome"] == "Goal"]

# every shot: marker size scaled by xG, low-value shots faded
pitch.scatter(team_shots["location"].str[0], team_shots["location"].str[1],
              s=team_shots["shot_statsbomb_xg"] * 900 + 40,
              c="#b4532a", alpha=0.55, edgecolors="white", ax=ax, label="Shots")
# goals highlighted in green on top
pitch.scatter(goals["location"].str[0], goals["location"].str[1],
              s=goals["shot_statsbomb_xg"] * 900 + 40,
              c="#0f5132", edgecolors="white", ax=ax, label="Goals")

ax.set_title("Arsenal — shot map (size ∝ xG)", fontsize=14)
ax.legend(loc="lower left")
plt.show()

The same shot map shown schematically in 2.2 — here generated from real coordinates with mplsoccer.

The player radar

Radars (spider charts) are the standard way to show a player's statistical fingerprint against positional peers. Each spoke is a metric, scaled to a percentile so the shape — not the raw value — carries the meaning. The interactive chart below is the same idea rendered for the web.

Principles for football charts

• Percentiles beat raw numbers for comparison — "83rd percentile for progressive passes" is instantly meaningful; "6.2 per 90" is not until you know the distribution.
• Always show the reference — versus league average, peers, or the opponent. A lone number is not a story.
• Direction-encode the pitch — make clear which way a team attacks; mirror opponent data so both read left-to-right consistently.
• Less ink, more signal — a coach has thirty seconds. Strip chart-junk; highlight the one thing that matters.
• Colour with meaning and access — consistent team colours, outcome-based encodings, and palettes that survive colour-blindness and a projector.

3.4 Combining Video and Data Using Professional Tools

Numbers convince analysts; video convinces players. The decisive skill in a club is welding the two so a statistic becomes a coachable moment on screen.

A PPDA of 7 means nothing to a winger. Three clips of the opponent's front three springing the press — each tagged to the exact second the data flagged — means everything. The pipeline that connects them is the core of applied performance analysis:

The data-to-video loop in practice

1. Find the pattern in the data

   e.g. "we concede 0.6 xG per game from balls into the left half-space."

2. Filter to the moments

   Pull every event matching the pattern with its timestamp.

3. Pull the clips

   The tagging tool returns those exact video windows.

4. Build the package

   A short, sharp reel — ideally with on-screen overlays (the run, the space, the number) — for the unit meeting.

3.5 A Day in the Life of a Professional Football Data Scientist

If the analyst (1.4) lives on the match-week calendar, the data scientist lives on the model-and-pipeline calendar. The work is less about the next fixture and more about the systems everyone else depends on.

In small clubs one person wears both hats; in elite setups they are distinct teams. Both need football understanding — a data scientist who can't tell a meaningful feature from a noisy one builds elegant, useless models.

4.1 Introduction to Use Cases

Everything so far — data types, metrics, tools — exists to serve a decision. Part 4 walks four realistic jobs an analyst is actually handed, each ending in something a coach or director can use.

The four cases below are deliberately the everyday core of the role, not exotic projects. They are illustrated with representative (not real) numbers so you can follow the reasoning, which is the transferable part. Notice that each follows the analytical loop from 1.2: a question, a hypothesis, evidence weighed against a baseline, and a clearly communicated answer with its uncertainty.

4.2 Reading Your Opponent: A Practical Case Study

The brief: "We play Riverside FC on Saturday. Tell the staff how they play and how we beat them." You have their last eight matches of event and tracking data.

Step 1 — Establish identity (the baseline)

Start broad. Over eight games, Riverside average 54% possession, a PPDA of 9.2 (aggressive press), and a 4-3-3 that becomes a 2-3-5 in possession with both full-backs high. Already a picture forms: a proactive, high-pressing, possession side that commits numbers forward. The tactical opportunity with such teams is usually the space they vacate behind the full-backs.

Step 2 — Find the weakness (test the hypothesis)

Hypothesis: their high press leaves them open to counters in behind. The evidence backs it — Riverside concede 1.6 xG per game despite dominating the ball, and 63% of that xG comes from transitions, most of it down their left where the full-back pushes highest. The defensive-action map confirms their line sits high; the tracking data shows their recovery runs are slow on the left side.

Step 3 — Pressing triggers & build-up

When we have the ball, where will they press? The data shows Riverside trigger their press on our centre-backs' first touch and aggressively jump our full-backs. Their weakness: the single pivot leaves the space between their lines open if we can play through the first wave. Plan: bait the press, then break it with a forward pass into the No. 10 dropping between their midfield and defence.

Step 4 — Set pieces & individuals

Granularity wins matches. Riverside concede a high share of their chances from near-post corners and zonal-mark the six-yard box — a routine to exploit. Individually: their left-back (the one bombing forward) is beaten 1v1 more than any defender in the division; their keeper's PSxG–GA is negative (under-performing his shot-stopping). Both go on the dossier.

4.3 Post-match Analysis: A Practical Case Study

The brief: "We drew 1–1 with a side we should beat. The coach is frustrated. Was the performance actually bad?" This is where analytics most directly protects a club from over-reacting to a result.

Separate process from outcome

The scoreline says "dropped points." The xG story may say something completely different. Suppose the match xG was 2.3 – 0.7 in our favour: we created chances worth more than two goals, conceded almost nothing, and drew because of one wonder-strike against the run of play and a string of saved efforts. That is a good performance with a bad result — and the correct message to the dressing room is "keep doing this," not "tear it up."

The post-match checklist

• xG for & against — did we deserve more? Were chances high quality or did we inflate xG with volume of poor shots?
• The shape of chance creation — open play vs set pieces, sustained vs one-off. A flattering xG built on one penalty is not the same as a steady stream of box entries.
• Finishing vs creation — was the problem getting into good areas (a tactical issue) or converting from them (a finishing issue, often variance)?
• Defensive concession — was the goal we conceded a systemic breakdown or a freak? One tells you to fix something; the other tells you to do nothing.
• Individual underlying numbers — who progressed the ball, who created, who lost it in dangerous areas.

4.4 Player Analysis: A Practical Case Study

The brief: "Our deep-lying playmaker is leaving. Here's a shortlist of three replacements within budget. Who fits, and what are the risks?"

Step 1 — Define the role, not the position

"Replace the No. 6" is too vague. From 2.5, define the role by function: this player receives under pressure, progresses the ball through passes and carries, dictates tempo, and screens the defence. That becomes a target statistical profile — high progressive passes and pass completion under pressure, strong xT generated, solid (not elite) defensive volume.

Step 2 — Compare like-for-like, in context

Now the discipline of 1.2 bites. Candidate B has gaudy progression numbers — but he plays for a dominant side in a weaker league, against deep blocks, with time on the ball he won't get here. Adjust for level of competition, team style and minutes before comparing. Percentile radars against positional peers in comparable leagues make the styles legible at a glance.

Step 3 — Weigh fit, quality, risk and value

4.5 From Data to Tactics: Building a Tactical Report

The tactical report is where the whole handbook converges: data types, metrics, tools and interpretation, compressed into a document that changes how a team plays on Saturday.

A report nobody reads is a failure regardless of its analytical quality. The craft is ruthless prioritisation — surfacing the two or three things that will decide the match and burying everything else in an appendix. Structure beats volume.

Anatomy of a tactical report

From insight to instruction

The final translation is the one that matters most. A data insight ("they concede 63% of their xG in transition") is useless to a player until it becomes an instruction ("win the ball and get it forward within three seconds — look for the run in behind their left-back"). The best analysts live in this last mile: turning a probability into a behaviour.

"The report is not the deliverable. The changed decision on the pitch is the deliverable. Everything else is plumbing."The closing principle of football data analysis

Glossary & Sources

A quick-reference glossary of the metrics and terms used in this handbook, followed by the sources consulted.

Core metrics & terms

Sources & further reading