Methodology

Definitions describe specific data contracts, not a guarantee that every deployed view shares the same filters. See eligibility by source. Historical examples and diagnostics must not be read as current coverage.

How much to trust what

Not everything here is validated the same way, and pretending otherwise is how you get caught out. Four tiers, strongest first.

Automated checks

Distinguish two different things on this page. The Tier 1 reconciliations and the Tier 2 calibration are automated invariants: they re-run on every rebuild, in a verification step that runs last and fails loudly by name. The Tier 3 reference benchmark is a one-time external comparison, run once by hand against published figures and retained as a record. It is not recomputed and would not catch a regression.

The automated checks live as rows in an invariants table, each holding SQL that returns a count of violations. The verification step raises and names every failing check if any error-level check is non-zero.

Adding a check is an insert rather than a code change, which is the only way a practice like this survives contact with a busy season. The current count is read live.

-- example: the goal reconciliation, as it actually runs
select count(*) from matches m
left join (parsed goals per game) ev on ev.game_id = m.game_id
where m.home_score <> coalesce(ev.h,0)
   or m.away_score <> coalesce(ev.a,0)

Other error checks: no team appears outside its league whitelist, every sequence carries the league of its own events, percentiles fall between 0 and 100, no sequence has a null threat value, the sequence layer covers exactly the games with events, and season xG stays within 10% of goals scored.

Pitch coordinates

The feed supplies coordinates on a 0–100 scale in both axes, always from the attacking team’s perspective, left to right. Converting to metres uses a 105 × 68 pitch:

x_metres = x / 100 * 105
y_metres = y / 100 * 68

Thirds are cut at x < 33.3, 33.3–66.7 and ≥ 66.7. The penalty area is x ≥ 83 with y between 21.1 and 78.9. Wide channels are y < 21.1 or y > 78.9.

Ingestion

Event data is scraped from WhoScored using the soccerdata library, one match at a time, with 60–120 second gaps. Only fixtures not already stored are fetched, so a repeated run costs nothing.

Write-time guard

Before any match is written, both clubs are checked against a per-league whitelist. A match containing an unrecognised club is refused rather than written. The guard fails open only when a league has no whitelist yet, so a new competition can bootstrap from its first scrape, and closed in every other case.

This exists because it did not, once. A scraper defaulting to a different competition wrote another league’s data into the database, and cleaning it up had to be done by fixture id rather than club name, because schedule names and event-feed names differ for more than half the clubs.

Expected goals

An empirical binned model fitted on this season’s shots. No external xG is used.

Features

distance = sqrt( ((100 - x) * 1.05)^2 + ((50 - y) * 0.68)^2 )

angle    = degrees( atan( 7.32 * ((100 - x) * 1.05)
                    / ( ((100-x)*1.05)^2 + ((y-50)*0.68)^2 - 3.66^2 ) ) )

7.32m is the goal width and 3.66m the half-width, so the angle is the true horizontal aperture the shooter sees.

Bins

Live shot xG uses a smoothed bin estimate; penalties use 0.76. Player xG per 90 includes penalties unless explicitly labelled non-penalty. Open-play pitch filters and historical saved model features have separate scopes. Do not compare them as identical totals; see Evidence definitions for fitted-version and snapshot caveats.

Predictive check. A live temporal holdout now fits the same feature family on the first 80% of current-season matches and scores it on the final 20%. Sparse cells fall back to broader training-only shot-shape rates. This is an out-of-time baseline, not external validation or a claim that the production lookup has been replaced.

Expected threat

The pitch is divided into a 12 × 8 grid. Each cell carries a value: the probability a possession in that cell ends in a goal. Moving the ball adds the difference.

xT(action) = grid_value(end_x, end_y) - grid_value(start_x, start_y)

x_bin = floor(x / 100 * 12)   clamped 0..11
y_bin = floor(y / 100 *  8)   clamped 0..7

Applied to completed open-play passes and to derived carries, reported separately as threat from passing and threat from carrying. A backwards pass produces a negative value and is counted as such.

The database publishes a separate internal directional check comparing xT on shot-ending and other possessions. It is explicitly a software sanity check; the live status retains externally_validated = false.

Possession sequences

A sequence is an unbroken spell of control by one team. Segmentation runs over events ordered by period, minute, second and event id.

Which events count as control

Pass, TakeOn, BallTouch, MissedShots, SavedShot, ShotOnPost, Goal, KeeperPickup, Claim

A new sequence begins when

• the controlling team changes,
• the period changes,
• a stoppage event occurs (Foul, Card, OffsideGiven, OffsidePass, CornerAwarded, End, or a goal),
• or the event carries a set-piece qualifier (ThrowIn, CornerTaken, FreekickTaken, GoalKick, KickOff, Penalty).

Set-piece-initiated sequences are flagged and excluded from every open-play figure.

Tags applied to each sequence

Chain roles

Each player’s involvements are mapped back onto the sequence they belong to. Eleven behaviours are expressed as a percentage of his total involvements. Goalkeepers are excluded, and a player needs at least 120 involvements to be profiled.

Tempo is recorded at whole-second resolution in the feed, so it is coarse for any single action and only meaningful as an average over many.

Chain-position value

early involvement = touch in a shot-ending sequence,
                    at least 3 actions before the end

Reported per 90. It credits the pass that starts a move rather than the one that finishes it. It is contaminated by teammate quality: a player alongside good forwards appears in more shot-ending sequences regardless of his own contribution.

Percentiles

A percentile here is arithmetic, not a model. There is nothing to validate; the question is whether the metric feeding it is right.

pct = round( 100 * percent_rank()
             over (partition by league, pool, metric order by value) )

-- inverted for metrics where lower is better, e.g. dispossessions
pct = 100 - pct

• Within position pool (CB, FB, CM, AM, W, ST), because distributions differ enormously by position.
• Within league, so a 60th-percentile progressor in one competition is not silently equated with another.
• Eligibility depends on the source. The committed mv_player_percentiles definition uses six nineties; the distinct mv_player_pct definition uses three. Frontend cohort settings are not proof of the deployed SQL gate. Chain roles and team insights have separate exposure rules. See the source contracts.

So “80th percentile finisher” means precisely: his finishing rate is higher than 80% of eligible players in his position pool in his league. It is a statement about ranking, not about quality in the abstract.

Pooling several leagues is offered separately. Pooling is not translation. It answers where a player sits among a chosen set, not how he would perform elsewhere.

Team metrics

Press profile

Every possession has an opponent, so from the defending side it is a pressing test. Opponent possessions are grouped by build-up type (short build, direct, mid, high start), and containment is measured as the share ending below x = 66.7 without a shot.

Raw containment is dominated by mechanics rather than quality, since direct play travels further by definition, so figures are z-scored within build-up type against the league. A first version of this reported a 50-point gap for every team, which was measuring the physics of long balls rather than anything about defending.

Game state

A goal timeline is built per match, with own goals credited to the correct side, and every possession is tagged with the score at the moment it began. Output can then be weighted by whether the game was still live:

margin within 1 goal -> weight 1.00
margin of 2 goals    -> weight 0.60
margin of 3+         -> weight 0.35

Both the output and the minutes are weighted, so this measures rate per live minute rather than penalising players at dominant clubs.

Usage and leverage

Squad role uses two axes rather than raw minutes.

selection share = minutes played
                / minutes available while at that club

leverage        = share of his on-pitch minutes
                  with the score within one goal

The availability window runs from a player’s first appearance for the club to the club’s most recent fixture, so a mid-season signing is not punished for games he could not have played.

Similarity

Players are compared on percentile vectors within position pool. Percentiles rather than raw values, because they are robust to the skew in most role distributions.

distance = sqrt( sum( (pct_a - pct_b)^2 ) ) / sqrt( n_shared )
similarity % = 100 - distance

The default comparison uses 32 dimensions spanning roles, pass trajectory, carrying, creation, shooting and defending. A narrowed query, such as similarity as a shooter, restricts the vector to that family and requires at least 60% of the target’s available dimensions.

Similarity is stylistic, not qualitative. The closest match to a good player is frequently someone who does the same things less well. And a single-metric comparison will return very high percentages simply because there is only one axis on which to differ.

Definitions versus measurements

This is the distinction that matters most when someone challenges a number.

Answering the fourth question as though it were the first is the fastest way to lose a room. Answering it precisely is the fastest way to earn one.

Known limits

• No off-ball information. The feed records the player on the ball. A run that drags a defender out of position is invisible. This is the largest limitation and it does not go away.
• No defender positions. Pass classifications describe trajectory, not verified line-breaking, because we cannot see who was between the two points.
• Single season. Team style is firm at this sample; individual finishing and anything rate-based on low volume is not.
• Age is a snapshot at match time, not a date of birth, so it can read a year light. The highest age observed is stored.
• 25 shots carry no coordinates and are absent from the shot model.
• Sequence tags are untuned against video.