The Football Language Model: 2,000+ Metrics That Describe the Modern Game

The original football statistics. Counted from 1888 onwards in match reports, formalised by Opta in 1996:

• Goals, assists, shots, shots on target.
• Passes attempted + completed, pass accuracy.
• Tackles, interceptions, clearances, blocks.
• Corners, free kicks, penalties.
• Yellow cards, red cards, fouls committed/won.
• Possession % (in three different methods).

Probability-based metrics built on Tier 1 counts. The breakthrough was xG:

• xG (expected goals). Weight every shot by the probability of scoring given location, body part, etc.
• xGOT (post-shot xG). Conditional on the shot being on target — measures finishing skill.
• xA (expected assists). Weight every chance-creating action by the xG of the chance produced.
• xGA (expected goals against). Defensive equivalent — quality of chances conceded.
• xPts (expected points). Probability of points won given xG vs xGA in each match.

Metrics that weight actions by their game-state importance:

• OBV (On-Ball Value). Every action rated by its contribution to scoring or conceding probability — leverage-weighted.
• xT (expected threat). Position-based threat scoring; pioneered by Karun Singh.
• Win-probability uplift. A goal in a 0-0 game is worth ~0.2 win-probability points; the same goal in a 4-0 game is worth ~0.005. Adjusts for game state.
• Adjusted defensive value. Tackles + interceptions + recoveries weighted by zone and game context.
• xT differential per 90. A combined attack + defence rating in xT terms.

Metrics that require positional data on every player at every moment, not just on-ball events:

• Pressure intensity. Per-second proximity-and-velocity-based pressure scores.
• Off-ball run quality. Quality of runs into space, regardless of whether the ball arrives.
• Defensive line height. Continuous tracking of where the back-line sits.
• Ball-recovery time. Average seconds from possession loss to win-back.
• Pass into pressure. Probability that a pass arrives at a pressed teammate.

Three drivers of the metric explosion:

• Better data infrastructure. Tracking systems and event-coding scale make it possible to compute metrics that were previously inconceivable.
• Recruitment-decision pressure. Multi-million-pound transfer decisions create demand for ever-more-specific metrics to differentiate similar players.
• Academic + industry research. Researchers publish new metrics; data providers commercialise them; clubs adopt the proven ones; the cycle repeats.

Three principles for using football metrics well:

• Start with the question, not the metric. "Is this striker a good finisher?" → xG vs goals. Don't start by browsing metrics looking for something to say.
• Pair Tier 1 with higher tiers. Goals + xG together is more informative than either alone. Tackles + interceptions + adjusted defensive value reveals what raw tackle counts hide.
• Trust the trend, not the spike. Single-match metrics are noisy. Five-match or season-aggregate metrics are signal.

Three predictions for the next 5 years:

• Causal-inference metrics. Moving from correlation ("teams with this metric tend to win") to causation ("doing X causes a Y change in win probability").
• Player-style fingerprints. Multi-metric vector representations of each player (similar to the player-profile widget approach), enabling similarity search at scale.
• Live in-match probabilistic updates. Real-time updating of metrics during matches, surfaced in broadcast graphics.