Football Analytics: xG & Beyond

Reep's notebooks became football's earliest data corpus. His analytical conclusions were later heavily criticised, but the practice of systematic event-by-event recording began with him.

Source: Coaching history / analytics literature · recall

Hughes wrote 'The Winning Formula' in 1990 formalising Reep's analysis into English FA coaching doctrine, prescribing direct play from POMO zones.

Source: FA coaching literature · recall

Opta Sportsdata began operating in 1996 and rapidly became the reference feed for English football event data before spreading to broadcast and betting markets.

Source: Opta historical archives · recall

Prozone's 1998 launch, initially adopted by Derby County under Jim Smith, established the template for club-level video-plus-data analysis at scale.

Source: Coaching history · recall

Wyscout's Italian launch in 2004 grew into the tool that virtually every European club uses for scouting reports, opponent analysis and player video libraries.

Source: Coaching history · recall

Benham bought Brentford in 2012 and Midtjylland in 2014, applying odds-style statistical inference to recruitment and match preparation at both clubs.

Source: Business Insider / Financial Times profiles · recall

Graham was Director of Research at Liverpool from 2012 to 2023, presiding over their most successful period since the 1980s: the 2019 Champions League and the 2020 Premier League among six major trophies.

Source: MIT Sloan Sports Analytics / Ludonautics · recall

Knutson founded StatsBomb as an analytics blog in 2013, launched their own event data in 2018, and remained CEO until Hudl acquired the company in August 2024.

Source: StatsBomb archives / Business Wire · recall

Rudd's 2011 paper defined the possession-state framework using transition probabilities. Karun Singh later popularised the specific 'Expected Threat' name in a 2018 blog post.

Source: Academic and industry analytics history · recall

Singh's blog laid out xT as a value surface across pitch zones, computed via iterative solving of transition probabilities. The post remains the industry reference introduction to the metric.

Source: karun.in / analytics literature 2018 · recall

xG models fit historical shot outcomes to features such as distance, angle, body part and pattern of play; more advanced models incorporate defender positions from tracking data.

Source: Opta / StatsBomb documentation · recall

PSxG models the probability a goalkeeper concedes given the on-target trajectory of the shot; it uses everything xG uses plus where the shot ended up going.

Source: StatsBomb documentation · recall

xA credits the passer with the xG of the resulting shot, so a great pass into a great shooting position is rewarded even when the finish is missed.

Source: Opta documentation · recall

Singh's xT computes each pitch cell's threat as the probability a team goes on to score within a small number of subsequent actions, then credits players who move the ball to higher-value cells.

Source: karun.in / xT literature · recall

OBV is StatsBomb's proprietary possession-value framework that assigns a value to every event including defensive interceptions and shots, addressing an xT limitation.

Source: StatsBomb documentation · recall

VAEP frames every event as a change in the probability of scoring or conceding in the near future, and assigns credit accordingly.

Source: KU Leuven Machine Learning Research Group · recall

Field tilt captures territorial dominance by measuring where possession is happening, not just how much of it a team enjoys — a team can have 60 per cent possession but 40 per cent field tilt.

Source: Analytics literature · recall

Packing counts the opposition outfield players the pass or carry bypasses; it was popularised by Germany's DFB analytics during the 2014 World Cup cycle.

Source: DFB technical reports · recall

The 360 snapshot fills in where every off-ball player was at the moment of each on-ball event, unlocking analysis of defensive shape and passing options that pure event data cannot see.

Source: StatsBomb 360 product docs · recall

PPDA is the classic press proxy computable from vanilla event data alone; OBV and 360-derived metrics need proprietary or tracking data that pure Opta feeds do not carry.

Source: Analytics literature · applied

Edwards was Liverpool's SD 2016-2022 during the era Graham described as 'the Premier League's first data science department', then returned to a wider FSG role in 2024.

Source: Athletic / Ludonautics profiles · recall

Comolli's short Liverpool spell was controversial in its immediate results but planted the seed for a data-oriented recruitment framework Edwards later scaled.

Source: Athletic profiles · recall

Ankersen was a public face of the Benham era, becoming known for arguing that Championship match outcomes were subject to more variance than table position implied.

Source: Athletic / Guardian profiles · recall

Randall co-founded StatsBomb with Knutson; both departed following the Hudl acquisition announced in August 2024.

Source: World Soccer Talk / Business Wire 2024 · recall

Riddersholm led Midtjylland to their first title in 2014-15, leaning on Benham-model set-piece routines and analytical scouting that Knutson helped design.

Source: Superliga archives · recall

Graham founded Ludonautics in 2023 after leaving Liverpool. Michael Edwards was announced early on as a non-executive director.

Source: Research Live / Training Ground Guru 2023 · recall

Graham's memoir-cum-analysis was published in 2024 and is one of the most-cited public texts on club-level football data culture.

Source: Penguin Books Australia / Aevitas 2024 · recall

Decroos, Bransen, Van Haaren and Davis's paper introduced VAEP at KDD 2019, becoming the reference for possession-value modelling in academic sports analytics.

Source: KDD 2019 / KU Leuven ML Group · recall

German journalist Christoph Biermann's 'Football Hackers' (2019 English translation) chronicled the rise of analytics-led clubs from Midtjylland to Liverpool to Brentford.

Source: Publisher records / SportsBook of the Year 2019 · recall

The pattern of Benham, Knutson and Graham is unmistakable: quant-heavy backgrounds outside professional football itself have consistently been the highest-leverage first hire for analytics-led clubs.

Source: Coaching trends 2013-26 · applied

A large negative gap between goals and xG usually corrects toward xG over larger samples; the base assumption is finishing variance, not model failure.

Source: Analytics literature · applied

PSxG minus goals is the standard shot-stopping proxy; -7 means 7 fewer goals than the on-target sample predicted, a strong positive signal on the keeper's shot-stopping.

Source: StatsBomb analytics · applied

A shot-map concentrated in the six-yard zone indicates a striker whose value comes from finishing service, not from creating shots at distance or wide angles.

Source: StatsBomb / Wyscout visualisation · applied

Progressive passes count metres of forward progress; xT rewards moving the ball into higher-value zones. A big progressive-pass volume with modest xT means the metres are being added in relatively safe areas.

Source: Analytics literature · applied

High field tilt with modest xG is the signature of late-game territorial pressure without penetration — a classic sign of a low-block opponent absorbing well.

Source: Analytics literature · applied

High attacking OBV with negative defensive OBV is the archetype of the attacking full-back — think Alphonso Davies or Trent Alexander-Arnold profiles.

Source: StatsBomb OBV literature · applied

Set-piece xG of 8 with only 4 goals is a plausible variance gap that regression can close; a review of execution is the low-cost intervention before wholesale change.

Source: Coaching literature · applied

Location matters more than volume; deep dribbles are lower-value events, and the map suggests role deployment rather than skill absence.

Source: Analytics literature · applied

Half-space receipts and elite carry contribution are the fingerprint of the inverted winger role, a profile Arsenal and City have specifically recruited for since 2022.

Source: Recruitment literature 2022-25 · applied

Off-ball threat only shows up when we know where the player was standing without the ball. Location snapshots at chance moments are the specific evidence needed.

Source: StatsBomb 360 use-cases · applied

Midtjylland's 2014-15 title was the first in club history and became the flagship case for public analytics writing about Benham's methodology.

Source: Danish Superliga records / Football Hackers · recall

Brentford's 2020-21 Championship playoff win gave the club top-flight status for the first time in 74 years, widely credited as a triumph of the Benham analytical model.

Source: Championship records 2020-21 · recall

Leicester's title season is the modern archetype for models being caught out by low-frequency, high-variance outcomes — precisely the risk profile Benham's approach was designed for.

Source: Premier League records / analytics literature · recall

Graham's public reflections describe Salah as the model-driven signing that most publicly validated the department. He arrived from Roma for around £36.9m and won every major trophy at Liverpool.

Source: Athletic / Ludonautics profiles · recall

Robertson's signing at £8m from a relegated Hull side is a canonical data-scout case: underlying numbers massively out of step with market perception.

Source: Athletic / Ludonautics profiles · recall

Liverpool's 2021 deal was the launch case for StatsBomb 360, which adds location snapshots for players around each on-ball event.

Source: Sky Sports / Business Wire 2021 · recall

Haaland's 36 Premier League goals in 2022-23 — the highest single-season total in the 38-game era — sat on top of xG per 90 that itself was elite, an unusual combination.

Source: Premier League records 2022-23 · recall

Brighton's De Zerbi era combined an analytics-first recruitment model with a distinctive build-up shape, and became a public reference for progressive-pass profiling.

Source: Premier League analysis 2022-24 · recall

Iraola's Bournemouth adopted an aggressive pressing model and a data-heavy recruitment approach that carried the club to Europa League qualification in 2025-26.

Source: Premier League 2023-26 · recall

Leeds's data-informed operation grew visible under Marcelo Bielsa and became a public analytics story under the 49ers, particularly in dead-ball and goal-kick modelling.

Source: Championship / Premier League analysis 2020-25 · recall

Public xG models converge on around 0.76 for a penalty: roughly three in every four are scored.

Source: Opta / StatsBomb model documentation · recall

Central six-yard-box open-play shots are among the highest-xG open-play situations, though still short of a penalty. The specific figure varies by model but sits in the mid-40s.

Source: Opta / StatsBomb model documentation · recall

Long-range central shots have very low xG despite their visual drama: even the very best strikers score them at a low percentage.

Source: Opta / StatsBomb model documentation · recall

StatsBomb documented its feed as 3,300 to 3,400+ events per match, adding location snapshots for every event under the 360 product.

Source: StatsBomb documentation 2021-24 · recall

Singh's canonical implementation used a 12x8 grid; implementations since have used up to 20x15 cells, but 12x8 remains the reference.

Source: karun.in blog 2018 · recall

The Robertson deal is treated as an emblematic case of data-scouting bargains: an eventual world-class full-back for a fee later joked to be one of the great steals of the era.

Source: Athletic transfer records 2017 · recall

The industry rule of thumb — corroborated in multiple Opta and StatsBomb season reports — is that dead balls account for roughly a quarter to a third of goals in most modern leagues.

Source: Opta / StatsBomb season summaries · recall

xT rewards destination-zone value; a pass into the penalty area is worth vastly more than a pass into the centre circle because scoring probability in the next few actions is much higher from box zones.

Source: karun.in blog / xT literature · recall

Under-performance versus underlying xG almost always corrects toward xG in larger samples; the intervention is patience and monitoring, not tactical overhaul.

Source: Analytics literature · applied

The bottleneck is described specifically: the team creates threat but does not convert it into shots or shots into goals. A finisher directly addresses the missing step from threat to xG.

Source: Coaching literature · applied

PSxG conditions on the on-target trajectory of the shot, isolating the keeper's response from finishing quality — the correct instrument for shot-stopping.

Source: StatsBomb documentation · applied

xG chain only rewards actions when a shot follows. xT rewards any ball-moving action into higher-value zones regardless of whether the possession ends in a shot — closer to the recruitment brief.

Source: karun.in blog / xG chain literature · applied

Off-ball movement is by definition not an on-ball event, so any measurement of it requires location data at times where the striker is not on the ball.

Source: Analytics literature · applied

Adding defender positioning tightens the model's predictions: a shot from the same location with a defender blocking the shooter has meaningfully lower scoring probability.

Source: StatsBomb / Opta model literature · applied

Single-match xG is a small-sample quantity; ranked against a season total, the match-level number is a much noisier statistic and headline conclusions from a single match rarely hold up.

Source: Analytics literature · applied

OBV values shots and defensive actions too, and a proprietary feed comes with quality assurance and coverage; hobbyist xT is a useful cheap start but is not a full recruitment stack.

Source: StatsBomb OBV documentation · applied

Simple shots have simple features on which both models agree. The disagreement grows in complex phases where each provider's contextual features diverge.

Source: Analytics literature · recall

Frequency maps show volume; xG maps show volume weighted by quality of position. The latter directly answers the question posed.

Source: Analytics literature · applied

Off-ball distance during opposition possessions is by definition not an event; it can only be reconstructed from continuous player-location tracking.

Source: Analytics literature · recall

VAEP is fully documented and reproducible from KDD 2019; OBV runs at scale on the vendor's live data pipeline. Each has a place, but the trade-off is transparency versus operational convenience.

Source: KDD 2019 / StatsBomb documentation · applied

Hudl's acquisition closed in August 2024. Ted Knutson and Charlotte Randall stepped aside; the combined company retained the StatsBomb name inside Hudl's product suite.

Source: World Soccer Talk / Business Wire 2024 · recall

Ludonautics was launched in 2023; Edwards's role was announced at launch, and Graham's book followed in 2024.

Source: Training Ground Guru 2023 · recall

Edwards returned to FSG in a wider football operations role in 2024, having stepped down as SD in 2022.

Source: Athletic reports 2024 · recall

Graham's 2024 book is centrally structured around the Liverpool era that peaked in the 2019-20 Premier League title, the club's first for 30 years.

Source: Penguin Books 2024 · recall

Alonso's Leverkusen won the club's first ever Bundesliga title in 2023-24, doing so unbeaten and adding a DFB-Pokal, cited across analytics literature as a case of coaching-plus-data working together.

Source: Bundesliga records 2023-24 · recall

Slot's Liverpool won the Premier League in 2024-25, delivering the club its first title since 2019-20 in his first campaign in charge.

Source: Premier League records 2024-25 · recall

The 2023-26 pattern across top clubs is that tracking data plus event data — StatsBomb 360, Second Spectrum, Sportlogiq — is the base layer; everything else sits on top.

Source: Analytics adoption trends 2023-26 · applied

Any leading indicator claim must be checked against variance: middle-third xT can be a noisy per-90 metric if the sample is small. The sanity check is a stability analysis, not a credential check.

Source: Analytics literature · applied

Claim validation on set-piece xG requires the specific metric across time, adjusted for schedule difficulty; single-season headline figures without adjustment are usually not decisive.

Source: Analytics literature · applied

xG faced of 4 in a single match indicates a lot of dangerous shots got through; the shutout is real but the underlying signal is that repeated defending of that quality of shot volume will concede.

Source: Analytics literature · applied

Finishing skill emerges only across very large shot samples; 10 shots is dominated by variance. The mistake is not the model, it is the sample size.

Source: Analytics literature · applied

A player on a team that concedes more time in the defensive third has more opportunities to tackle. Absolute tackle counts confound skill with exposure; the fix is a rate-per-opportunity adjustment.

Source: Analytics literature · applied

Four seasons of consistent 25 per cent over-performance is enough evidence to update beliefs about the player's finishing. Insisting on regression at that horizon is a distinct kind of error.

Source: Analytics literature · applied

One loan spell in a specific division is a small, non-representative sample; using it as a decisive input is a textbook overfitting risk in recruitment.

Source: Analytics literature · applied

OBV can be strongly negative for a player whose role forces high-risk actions; the check is to compare his OBV to positional benchmarks and account for tactical instructions.

Source: StatsBomb OBV documentation · applied

Without defender context, a crowded six-yard-box shot with three defenders on the line is treated the same as an unmarked shot. The model over-values crowded set-piece shots by design.

Source: Analytics literature · applied

xT values are learned from historical data. If the modern game weights the half-space more heavily than the 2017-18 sample did, the reused model under-weights that zone.

Source: xT literature · applied

Clutch analysis is a classic base-rate error waiting to happen: last-minute shots have their own base rate. The specific fix is comparing to the same player's non-clutch performance and to positional baselines.

Source: Analytics literature · applied

Model risk in a single-vendor stack is a specific concentration risk: every model has assumptions, and betting the operation on one vendor bakes in its biases wholesale.

Source: Risk management literature · applied

Aerial-duel win rates per match are famously noisy; single-match splits are usually dominated by variance, and a projection based on that split is a textbook small-sample error.

Source: Analytics literature · applied

xG is a probability, not a guarantee. An 0.15 shot scores about 15 per cent of the time across enough attempts; that is the entire content of the number.

Source: Opta / StatsBomb documentation · recall

The core motivation of xT is to credit valuable ball-moving actions that traditional statistics simply cannot see because no goal or assist follows.

Source: karun.in blog / analytics literature · recall

Liverpool's multi-year StatsBomb deal announced in 2021 was pinned to the StatsBomb 360 launch; the club used the tool alongside its own in-house department under Graham.

Source: Sky Sports / Business Wire 2021 · recall

The correct question is: which action has higher expected goal value? A short square to an unmarked shot from 12 yards typically dwarfs a weaker-foot 20-yard shot; the decision is straightforward xG maximisation.

Source: Analytics literature · applied

Sample-size checks are the single-highest-leverage sanity test on any data-driven scouting claim; a beautiful chart on 500 minutes tells a different story than the same chart on 5,000.

Source: Analytics literature · applied

The base rate of set-piece goals across the professional game is around 25-35 per cent regardless of style; the possession claim is orthogonal to that base rate.

Source: Opta / StatsBomb season summaries · applied

Aggregating xG and xA needs a per-90 denominator and a positional benchmark; without both, the compound score reflects role and playing time rather than isolated player quality.

Source: Recruitment analytics · applied

Match-level xG is a small-sample statistic that headlines but does not prove; the correct posture toward a single-match differential is provisional interest.

Source: Analytics literature · applied

The Benham-Knutson pairing at Brentford/Midtjylland is the emblematic data-first hiring of the modern era, later scaled at Liverpool via Edwards and Graham.

Source: Coaching / analytics history · recall

Public models are useful triage, but they lack defender context and off-ball structure that proprietary and tracking feeds provide — the specific reason recruitment operations pay for the upgrade.

Source: Analytics literature 2023-26 · applied