The Football Data Compendium: How Modern Analytics Revolutionized the Game

Modern football analysis rests on a small family of measurements that were almost unknown a generation ago. This compendium explains what each of them computes, where the models disagree, and why a career total without its counting rule is not a record at all.

Why football needed a new kind of number

For most of its history the sport was described by goals, appearances and league tables, three objects that answer what happened and say nothing about how. A side that created nine good chances and lost one-nil carried the same record as a side that created none.

The new metrics describe the process rather than the outcome. Expected goals values a chance before anyone converts it, pressing metrics describe where a team engages, and progression metrics describe how the ball travelled, naming the part of the game the scoreline compresses into silence.

Expected goals and how it is built

Expected goals is a probability model trained on a historical archive of attempts. Each shot is described by its features, chiefly distance to goal and the angle to the posts, and the model learns how often comparable attempts became goals before applying that relationship to a shot not yet taken.

The design is deliberately average: it asks how often a shot like this is converted by an average player, excluding the identity of the shooter on purpose. That choice is the source of most public confusion, because it describes chance quality and is not a verdict on a finisher.

The non-penalty variant and the value of specificity

Penalties are the most valuable and least informative standard event in the sport, worth about three-quarters of a goal regardless of the move that produced them. Removing them isolates open-play creation and sharply reduces the match-to-match volatility of a total.

The variant also removes a competition-specific constant, since leagues differ in how often penalties are awarded. The rules of honest reporting follow: non-penalty expected goals must be compared with non-penalty goals, and penalties belong on their own line rather than folded into a blended total.

Possession read as pressure

Possession share was the first advanced statistic and remains the least informative. It counts every touch equally, including harmless circulation in a team's own half, and a side behind only because the opponent is content to concede the ball. Field tilt corrects much of this by counting only attacking-third activity.

Pressing metrics go further and describe where a team chooses to engage. PPDA divides the passes an opponent completes in a defined zone by the defensive actions made in that zone, so a low value reads as frequent intervention. The zone must be stated, because a deep block barely registers inside a standard window.

Progression and the value of moving the ball

Progressive passes and carries count the actions that advance the ball meaningfully towards goal, and the distance threshold is a convention rather than a law. The metric describes territory gained and says nothing about danger created, which is why it is always paired with a value model.

Expected threat assigns a value to every position on the pitch, estimated by working backwards from goals through the chains of moves that produced them, and prices an action by the rise in that value. Needing only coordinates, it can be computed for leagues where tracking data has never existed.

The goalkeeper, measured properly

Goalkeeping was the last position to receive a serious metric, because shot volume is low and the events are rare. Post-shot expected goals solved the central problem by revaluing the attempt after it has been struck, adding placement and speed to the pre-shot geometry, so the keeper is judged on the shots that arrived.

Subtracting goals conceded from post-shot values produces goals prevented, expressed in the currency of the game. Its unit makes it attractive and over-read: season-to-season repeatability is weak, defensive shot profiles confound it, and one strong campaign is weak evidence about ability.

Why two credible totals disagree

Providers disagree because they define events differently, not because one is careless. A blocked shot may or may not count as an attempt; a deflection may be attributed to shooter or defender; a cross that drifts goalwards may be logged as a delivery or a shot. Small numbers of events move a total.

The divergences are systematic rather than random, which is the important property. A pipeline that counts more tackles reports higher tackle rates for every side in the league, so the difference does not average out. Mixing providers inside one table is therefore a genuine error.

The sample-size discipline

Volumes stabilise quickly, value measures take longer and rates take longest, because they are built from the rarest events. Team expected goals needs most of a season to describe a level, a player's conversion rate needs multiple seasons, and an individual defensive metric grows slowly.

The practical rule is to name the sample in the sentence. Nine goals from a non-penalty expected goals of six over a full season is a claim that can be checked; he is overperforming is not, because the second version has deleted the evidence. Marking an unstable figure as unstable is itself a conclusion.

The records vault and the counting rule

An all-time record is a number plus a rule plus a date. Career scoring totals depend entirely on which matches are admitted, and two tallies of the same career can differ by several hundred goals without either being arithmetically wrong. Wartime championships, regional leagues and tour matches are the usual source.

This compendium states a rule before it states a total, reports a range where the classification is contested, dates its readings and describes nothing as official. No authority registered every competitive match in the sport, so no global ranking can be official.

The managers and the dynasties

A managerial trophy total is a count of whatever the compiler decided to count. National cup structures differ between countries, super-cups are major honours to one compiler and curtain-raisers to another, and comparable careers can differ by several places in a ranking that appears objective.

A dynasty is better described by metrics than by trophies, because metrics record the margin rather than the medal. The gap to the league average, the goal difference per season and the expected-goals differential express the size of an era's dominance, and they usually decline before the trophies stop arriving.

How to read everything that follows

Three habits make every figure on this site usable. Check the rule before comparing two numbers, because figures counted under different conventions are not in the same units. Check the sample, because a difference inside the noise band is not a difference. And check the date, because records and career totals are moving quantities.

Nothing in this compendium describes a fixture to come. Every metric measures play that has already happened, and every record is a reading at a stated date under a stated rule. That is a narrower promise than a prediction service makes, and the only one a reference work can keep.

Key reference points

• Expected goals is a learned probability over a historical shot archive, not a formula.
• The model is trained on an average player, so finishing skill is excluded by design.
• Possession share counts everything; field tilt and PPDA describe where a team engages.
• Progression measures territory; threat models measure danger. They are not the same.
• Post-shot expected goals is the only fair denominator for shot-stopping.
• Two credible totals differ because their definitions differ, not because one is wrong.
• Volumes stabilise fast, values slower, rates slowest of all: name the sample.
• A record is a number plus a counting rule plus a date, and none can be dropped.

A measurement is a tool with a stated purpose, a definition and a sample, and the purpose of this compendium is to keep all three attached to every number it prints.