Set-Piece Analysis - Why Restarts Dominate Modern Soccer Research

Spend 48 hours with any Premier League recruitment team and you will hear the same instruction: filter every target through corners and free-kicks first. Brentford’s 2026-24 accounts reveal £3.1 million-11 % of their total transfer pot-earmarked solely for players who add >0.08 expected threat per dead-ball action. The club climbed from 14th to 9th while shaving eleven seconds off the average time between whistle restart and shot. Copy the metric: log each candidate’s last 150 set-pieces, bin actions into six-second windows, keep only those who beat league median xT by 25 %.

The shift is measurable across Europe. Opta’s warehouse stores 1.04 million corner sequences since 2018; goals rose from 3.2 % to 4.6 % of them, but big-chance assists inside the subsequent eight seconds jumped 42 %. Arsenal exploited the spike, designing 42 pre-planned corner routines last season, scoring 14 goals-double their 2021-22 haul. Analysts now code five variables before the cross: blocker run angle, receiver’s shoulder orientation, keeper’s initial step, defensive line height, and second-post marking. Feed those into a gradient-boost model and you get a 0.71 AUC for predicting shot quality within three touches.

Clubs sell these edges for profit. Brighton bought 21-year-old Simon Adingra after data showed his inswingers created 0.17 xG per corner, top-5 in the Champions League. One season later, they loaned him to Toulouse with a €17 million purchase option, turning a €6 million fee into pure capital gain. The pattern repeats: identify under-23 wide players whose delivery speed averages >38 mph, buy below €10 million, insert sell-on clauses, and flip after one productive loan.

Track the next frontier: throw-ins. Danish club Midtjylland logged 1,800 long throws in 2025-26, generating 0.06 xG per sequence. Liverpool hired their coach, added GPS vests to measure hip rotation, and saw Trent Alexander-Arnold increase progressive distance per throw by 11 %. Within six months, the club’s post-throw scoring probability rose 0.004 per event-tiny, but worth three extra goals across a 38-match schedule.

Quantifying the 0.7%: Turning Throw-Ins Into xT Adders With Pass-Angle Vectors

Arm your right-back with a 28-32° reception arc and instruct him to stand exactly 2.4 m inside the touchline; Opta’s 2026-24 dataset shows this alone lifts expected threat (xT) from 0.007 to 0.019 per throw-in. Flat back-lines compress 1.2 m toward goal when the ball is re-entered from that angle, opening the half-space lane that leads 63 % of subsequent sequences into the final third.

• Target the front-post shoulder of the nearest centre-half; vectors drawn from 1 423 successful receptions reveal a 0.38 xT spike when the ball arrives there inside 0.9 s.
• Demand a two-touch limit: touches > 2 collapse xT to 0.004, the 0.7 % club average.
• Pre-plan four shape rotations triggered by the linesman’s whistle; sides using these add 2.1 xT per match, roughly one extra high-value shot every 5.6 games.

Track the angle between passer, ball and next receiver; anything wider than 74° drops completion to 52 % and drags xT below the waterline. Liverpool’s data team filters for 55-65° sweet-zone arcs and hit 71 % retention, turning a dead phase into 0.26 xT accumulated across 38 fixtures-equal to one penalty gained.

Mid-table Championship clubs without bespoke schemes bleed 0.41 xT against per throw; simulate five basic patterns in Friends of Tracking code, pick the one that minimises opposition counter-angles, rehearse twice a week. Brentford added 8.3 expected goals difference season-on-season solely through these micro-margins.

Build a dashboard: input latitude/longitude of landing spot, whistle-to-release time, defensive line height; output is instant xT delta. Leicester’s U-23 squad used it for six weeks, saw 0.9 xT+ per 90, then promoted the module to first-team analysts. Copy the Git repo, calibrate to your league’s tracking framerate, and the 0.7 % stops being noise-it becomes scoreboard impact.

Corner Kick Clustering: k-Means on 17,000 Deliveries to Spot the Free-Header Zones

Set k=6, filter for inswingers landing between 6-12 m from goal, and mark the unclustered 4 % of the data as the ghost zone where markers lose their checks; clubs using this cut-off saw a 0.18 xGOT drop on second-ball shots compared with those who left the zone empty. The clusters expose a 1.3 m² pocket at (x=89.4, y=25.7) where attackers gained 0.41 s of aerial freedom; deliveries aimed 0.8 m deeper and 0.6 m closer to the six-yard line shift the expected header angle from 14° to 31°, raising conversion from 8 % to 19 % without changing the service speed.

Cluster ID	Mean x (m)	Mean y (m)	Att header time (s)	Def header time (s)	Free-header share (%)
1	88.9	21.4	0.72	0.53	42
2	90.7	28.9	0.68	0.61	17
3	87.1	25.1	0.91	0.55	63
4	89.4	25.7	0.93	0.52	71
5	90.2	23.0	0.69	0.64	12
6	88.0	30.2	0.65	0.59	9

Overlay man-marking heat maps: if the nearest defender starts ≥2.4 m away, the attacker wins 78 % of aerials; inside 1.9 m the rate collapses to 26 %. Teams releasing the blocker at 0.9 m from the keeper’s line and sprinting to cluster-3 coordinates increase unpressured headers by 11 per season; Brentford and Lens converted these into six extra goals in 2025-26. The model flags any corner with a >65 % ghost-zone probability; coaches receive a 15-frame clip showing the drag-back run that clears the lane, cutting preparation time from 45 min to 90 s.

Goal-Kick Press-Resistance Index: Measuring Seconds-to-Midfield vs. PPDA Drop

Clip every goal-kick at 15 fps, tag the frame the ball crosses 40 m from goal, stopwatch the elapsed time; sides below 3.2 s earn a +1 index point, above 5.1 s score −1. Liverpool 2025-26 averaged 2.9 s, City 3.4 s; both sit in the 90th percentile for press-resistance.

PPDA for the same sequence drops 18 % on average when the first pass travels 25 m or more; Brighton’s drop is only 7 % because Caicedo re-presses inside two touches. Add (PPDA_before − PPDA_after) × 0.15 to the index; values > 0.45 flag a team that breaks the press reliably.

• Index > 0.60 → target wide channels, force inward; concede throw-ins rather than short GK repeats.
• Index 0.20-0.60 → stagger midfield line 5 m deeper, trigger press at second bounce.
• Index < 0.00 → leave striker high, man-mark pivot, accept 55 % possession, attack transitions.

Arsenal versus Brentford, MD-8 2026: Ramsdale-to-Saliba-to-Rice took 2.1 s, index +0.72; Brentford PPDA fell from 9.4 to 6.8. Arsenal generated 0.37 xG within 20 s of the sequence, match-winning margin.

Goalkeeper velocity adds noise: a 0.3 s release advantage equals +0.08 index. Model release speed separately; subtract 0.02 for every 1 km/h above 42 km/h to keep index centred on build-up quality, not arm power.

Export the metric to training: run 5v4 high-block drills, award points only if midfield line is breached before 3.5 s; repeat until squad average index exceeds +0.50 across twenty kicks. Sides that hit the threshold pick up six extra points per season from restart-to-goal sequences.

Free-Kick Simulation Stack: Combining Wall Height, Ball Speed and Keeper Reaction for ExpG

Set wall height at 1.78 m-tracked height, not listed height-when the shot origin is 19-23 m central. Every extra centimetre trims keeper’s visible fraction by 0.8 %; Opta data show xG drops 0.012 per centimetre. Feed real-time skeleton data into Unity 3-D at 120 fps; ray-cast from ball to top of wall, flag first occlusion frame, freeze keeper’s head orientation vector, feed that vector into a 128-neuron LSTM trained on 14 832 historic penalties and 6 411 direct shots. Output: 0.23 s earlier reaction cue, cutting conversion by 4.7 %.

Ball speed distribution peaks at 26.4 m s⁻¹ for scored Premier League free kicks; anything below 23 m s⁻¹ lets elite keepers shift centre-of-mass 0.9 m before contact. Calibrate speed off foot-mounted IMU: 1 kHz accelerometer fused with 250 Hz optical marker, RMSE 0.18 m s⁻¹. Feed speed scalar into Navier-Stokes drag model with 0.29 surface roughness coefficient for latest Nike Flight; solver outputs 0.72 s flight time from 22 m. Merge with wall occlusion timestamp; if keeper reaction window <0.65 s, predicted save probability jumps from 17 % baseline to 38 %.

Wall density matters: track average player shoulder width, multiply by four, divide by goal width. Ratio >0.62 forces shooters into 1.8° tighter vertical angle; expected goal value falls 0.04. Collect width off high-speed contour footage, not roster data-padded shoulders add 3 cm per player. Update density live; feed into reinforcement learner that shifts keeper starting position 7 cm laterally per 0.01 ratio increase, pushing post coverage from 78 % to 83 %.

Combine three streams-wall height scalar, ball speed, keeper reaction window-into single tensor (3×1). Feed through shallow regression net with dropout 0.15; train on 2 600 labelled shots, achieve R² 0.81 on withheld Championship set. Output is shot-specific expected goal figure; typical range 0.06-0.34. Push value to bench tablet before ball is stationary; coaching staff see colour band: red <0.10, amber 0.10-0.20, green >0.20. Sides using the stack conceded 3.1 fewer goals from dead balls per season across 2025-26 EPL and Championship sample.

Post-match audit: sync high-speed cameras (300 fps) with IMU timestamp, back-calculate true wall height, speed, keeper first movement. If predicted xG deviates >0.025 from model, tag clip, append to training pool; nightly retrain cycle (7 500 iterations, Adam lr 1e-4) keeps drift below 0.008 across season. Ship updated weights to edge device (Nvidia Jetson Nano, 5 W) inside goalkeeper’s backpack; inference latency 11 ms, battery lasts 92 minutes-enough for warm-up plus first half. Clubs pay £1 200 per camera plus £180 monthly GPU fee; cost per goal prevented: £42 k, below current market price for one Championship point (£110 k).

FAQ:

Why do restarts matter more than open-play chances when both end up as goals?

Open-play goals look prettier on highlight reels, but restarts are the only part of the match a coach can script. Players know the exact spot of the ball, the defence is still, and the angle is fixed. Clubs have built entire models that treat every free-kick, corner, or throw-in as a mini-set-piece and then simulate thousands of variations overnight. The result: a routine that looks simple on television can be worth 0.25 expected goals before the ball is even kicked, something impossible to guarantee during the chaos of normal play.

How did Brentford and Midtjylland turn throw-ins into a scoring weapon?

They stopped treating throw-ins as throwaway restarts. Analysts clipped 20 000 clips, tagged where the first touch was received, and ran cluster analysis. One pattern stood out: if the throw-in is taken within 0.8 s and the first pass is played back toward the touch-line, the pressing team usually overshoots, leaving a half-space open. Brentford rehearsed this so often that the ball boys were instructed to give the player a dry towel only when the routine was on, shaving another 0.3 s. The payoff was 11 goals from throw-ins in two Championship seasons, roughly a quarter of their total.

What kind of data do clubs collect for corners that fans never see?

Every camera frame is converted into 29 tracking points. Analysts log the flight time, drop zone, whether the first contact is offensive or defensive, and the height of the attacker’s jump relative to his marker. Clubs then run kernel density plots to find the 1 m² patch on the six-yard box where the keeper is least likely to intercept. One Premier League side discovered that if the ball is whipped in at 38-42 mph and arrives between 8.2 m and 8.7 m from the goal-line, the header conversion rate jumps from 3 % to 11 %. That tiny band of grass is now nicknamed the cash zone on the training ground.

Can a poor defensive team fix its restart problems mid-season, or is it a long-term project?

Short corners can be patched in a week; zonal versus man-marking switches need a month. The limit is muscle memory. When Liverpool moved from a hybrid zonal setup to a full man-marking scheme in November 2020, they needed 11 training sessions before the error rate on aerial duels dropped. Video game-style VR headsets helped: players wore goggles that replayed the exact flight path while they jumped on a force-plate, timing the header 200 times a day. Over the next ten league games they conceded only one goal from 42 corners, down from six in the previous ten.

Is there a restart trend that has already peaked and is now fading?

The near-post flick at corners is quietly disappearing. Data departments noticed keepers now pre-move a step to that post, cutting the success rate from 9 % in 2016 to 3 % last season. Instead, clubs are punting the ball beyond the back post for a reverse header back across goal, a technique that jumped from 12 attempts in 2018-19 to 112 last year. The next fad looks like short kick-offs: three quick passes from the centre circle can move the ball 60 m while the opposition is still setting its press, and early adopters already average 0.04 expected goals per match from it—tiny, but it costs nothing.