How does Man City's 650,000-player database actually work?

What is a recruitment database, in plain terms?

It is a single structured record of every player a club might conceivably consider, held in one place so that they can be queried in the same way. Each entry carries identifying information — age, position, contract status, league, minutes played — alongside performance measures drawn from match data. The point is not that anyone reads 650,000 profiles. The point is that the population is defined in advance, so that a search can be run across all of it rather than across the small subset a department happens to already know about. A database of that size is closer to a census of the professional and semi-professional game than to a shortlist. Everything else is filtering.

Why hold that many players when a club signs only a few?

Because the cost of holding a player in a database is close to zero, and the cost of not having heard of the right one is high. Recruitment departments are trying to avoid a specific failure: the player who would have fitted, in a league nobody was watching, discovered only after a rival signs him. Breadth is insurance against that. It also changes the shape of the search. If the starting population is the whole game, the first question becomes "who meets these conditions?" rather than "what do we think of this name?" — which reverses the usual order, in which a name arrives first and analysis is used to justify or reject it afterwards.

Where does the underlying data come from?

Mostly from commercial providers who log matches in a standardised way. Event data records discrete actions — passes, carries, duels, shots, defensive interventions — each with a location, a time and an outcome. Tracking data, where it is available, records the position of every player and the ball many times a second, which allows off-ball movement and team shape to be measured rather than inferred. Coverage is uneven: top divisions are logged in detail, lower and more distant leagues far less consistently. Clubs supplement this with their own video work and their own derived models. Exactly which providers Manchester City use, and what they build in-house, is not public.

How can players in different leagues be compared fairly?

This is the central technical problem, and it is why raw numbers are close to useless on their own. A performance in a weak league is not equivalent to the same performance in a strong one, so models apply a league-strength adjustment, calibrated on players who have moved between competitions and on results in continental matches. The output is an estimate of what a player's contribution would translate to at a higher level. Age is treated similarly, through ageing curves that describe how particular qualities typically develop or decline. Both adjustments are estimates with uncertainty attached, and a well-built system reports that uncertainty rather than hiding it behind a single rating.

What does the data say about how a player actually plays?

More than whether he is good — it describes what he does. Clustering methods group players by the pattern of their actions: where they receive the ball, how far they carry it, whether they progress play through passing or dribbling, how high and how aggressively they defend. That produces role profiles rather than position labels, so two players listed as right-backs may sit in entirely different groups. For a club with a defined style, the profile matters as much as the rating. A side that builds from the back needs defenders comfortable receiving under pressure; a side that presses high needs forwards whose defensive work is structural, not occasional.

Does any of this replace scouts?

No, and no serious department claims it does. Data narrows an unmanageable population to something a human team can watch properly. Video and live scouting then answer what the numbers cannot: how a player behaves when the game turns against him, whether his positioning is instructed or instinctive, how he handles a system he is unfamiliar with. The two streams are meant to disagree sometimes, and the disagreement is informative. What changes under a data-led model is sequencing — analysts define who gets watched, scouts decide what watching reveals — and the discipline of having criteria written down before anyone forms an opinion.

How does a longlist become a shortlist?

Through successive filters, each one cheaper than the one after it. First come hard constraints: position, age band, contract situation, league, realistic availability. Then model-based screening on the qualities the club has specified. Then role and style fit against the way the team plays. Then human review of video, then live viewings, then medical, character and financial assessment. At each stage the population shrinks by an order of magnitude and the cost per player rises. The design principle is to spend expensive attention only on candidates who have already survived cheap tests — which is why the size of the initial database matters so little to the workload at the end.

How does the head coach's system shape the filter?

Decisively, because the criteria are derived from it. A model cannot rank players in the abstract; it ranks them against a specification, and the specification encodes what the team is trying to do. If a side wants to defend a high line, recovery speed is weighted heavily and aerial dominance less so. If it wants to control possession in the opposition half, the value of a midfielder's progressive passing rises and his ball-winning volume falls in importance. This creates a dependency: recruitment built tightly around one system becomes harder to repurpose if the system changes. Clubs manage it by recruiting for club-wide principles as well as for the current manager's preferences.

Why do rebuilds get concentrated into a short period?

Squads age in cohorts. Players signed together, at similar ages, tend to decline together, and contracts signed together expire together. A squad that has been successful for several seasons therefore tends to arrive at a point where several positions need renewing at once. The alternative — replacing one or two positions a year — is the theory, but competitive pressure, contract timing and market availability rarely cooperate. BBC Sport describes Manchester City's data department as a key part of a rebuild across the past 18 months, which in practice means several transfer windows treated as a single planning problem rather than as separate events.

What can a database not tell a club?

A good deal. It cannot price a player, which is set by a selling club's incentives. It cannot predict adaptation to a new country, language or level of scrutiny. It cannot capture how a signing changes the players already there. Its measurements are only as good as the coverage underneath them, which thins out quickly below the major leagues. And it is backward-looking by construction: it describes what has happened, under one set of team-mates and instructions, and asks a model to extrapolate. The honest version of data recruitment treats its outputs as structured evidence that narrows uncertainty, not as forecasts that remove it.

Sources and further reading

• BBC Sport — reporting from inside Manchester City's data department and its role in the recent squad rebuild.
• Commercial football data providers — public documentation on how event and tracking data are collected and standardised.
• Academic sports analytics literature — peer-reviewed work on league-strength adjustment, ageing curves and player similarity models.
• Club and league governing-body publications — official material on registration rules, contract terms and transfer window scheduling.