What xG misses about set pieces
Expected goals is the closest thing soccer has to a unifying statistic. It has been a quiet success — broadcasters use it, commentators are starting to, casual fans know roughly what it means. The model that produces xG, however, was built for open play. Set pieces, which account for somewhere between a quarter and a third of all goals scored in the top European leagues, are where the model struggles, and the gap between what xG thinks about set pieces and what the best teams have figured out about them is one of the more interesting open spaces in modern soccer analytics.
Why the model is fine on open play
Open-play xG works because open-play shots happen in a relatively predictable distribution of game states. A shot from twenty yards, central, off a through-ball with no defender between the shooter and the goalkeeper, has a fairly consistent conversion rate across many years and many leagues. The model can learn that rate, and once it knows it, it can apply it to any new shot with similar characteristics. There is variance, but the underlying distribution is stable enough that the model's predictions are useful in aggregate.
The trick is that "similar characteristics" is doing a lot of work in that sentence. Open play tends to produce shots whose determinants are well-captured by the standard xG features: location, angle, body part, assist type, whether it was a rebound, whether it was a header or a foot shot. Across millions of open-play shots, the model has enough data to smooth out the variation and produce a tight estimate.
What's different about set pieces
Three things, each one a real problem. The first is that the model features that work on open play don't capture what actually determines a set-piece shot's conversion rate. The location of a header off a corner is meaningful, but the more important variables are the routine the attacking team ran, the marking scheme the defense was in, and the specific contact the attacker made on the ball. Public xG models see approximately none of this. They see the location and the body part and apply an open-play conversion rate that is substantially different from the actual conversion rate of the shot in question.
The second is that set pieces are not independent events. The same team taking corners in the same match is running variations on a small number of pre-rehearsed routines, and the success of each routine is correlated with the others — if the defense has figured out the team's near-post pattern by the third corner, the fourth one will convert at a lower rate. xG models treat each shot as independent. The actual probability of a goal from a set piece depends on which previous set pieces have already happened in the same match.
The third is that the variance on set pieces is much higher than on open play. The same routine, executed identically, will convert wildly differently depending on whether the attacking team's central defender wins his individual duel with his marker. Soccer's individual-duel layer is, in general, hard for any spatial model to capture, and it's especially important on set pieces, where the bulk of the scoring chance is determined by who wins one specific battle in the box.
The Brentford and Brighton case
Brentford and Brighton, both with reputations for analytical sophistication, were unusually open about treating set pieces as a separate strategic problem from open play through the mid-2020s. Both clubs hired set-piece specialists — not assistant coaches who also handled set pieces, but dedicated staff whose primary job was set-piece design — and both produced consistently outsized set-piece returns relative to their open-play xG.
The numbers are striking. Across multiple Premier League seasons, both teams converted set-piece shots at rates meaningfully above the league average. Both teams' open-play xG-against and xG-for were unremarkable for their league position; their set-piece performance was the part of the ledger that pushed them above their open-play baseline. By public xG metrics, they were not exceptional teams. By actual table position, they were.
The Brentford approach in particular was studied within coaching circles: an aggressive use of the second-ball principle, several layered routines with multiple decision points based on defensive reaction, and a specific preference for short corners to draw defenders out before the cross. The set-piece routines themselves weren't secret — opponents could watch them on tape — but the defensive coordination required to fully cover them required dedicated preparation that most teams hadn't invested in.
Why most teams haven't caught up
Set-piece work has cultural problems within the sport. The traditional coaching hierarchy treats set pieces as a peripheral skill, something a coach handles with a whiteboard for twenty minutes before training and then moves on from. The teams that have made set pieces a strategic priority have generally done so by upending that hierarchy — putting an expert in the room with the head coach as a peer, not a subordinate, and giving him real time on the training ground.
Most clubs aren't structured to absorb that kind of reorganization. The head coach's authority over training priority is total in most setups, and a head coach whose instincts say set pieces matter less is going to allocate less time to them even if the analytics department wants more. The Brentford model — set-piece specialist with real sessions and real input — has been adopted by some other clubs but is still the minority approach.
What the next generation of xG might fix
Two paths. The first is set-piece-specific xG models that train on a much larger feature set: the formation of the attackers in the box, the marking scheme of the defenders, the routine type, the second-ball pattern. These models are starting to appear in academic papers and inside front offices, though public versions are still rare. They promise to fix the conversion-rate problem by recognizing that set-piece shots aren't drawn from the same distribution as open-play shots.
The second is sequence-aware models that track multiple set pieces in the same match and condition the conversion rate of each on the history of the others. This is harder because it requires modeling defensive adjustments over time, but the framework is the same as in-game adaptation models that already exist for serve patterns in tennis and pitch sequences in baseball. The sport's analytics community has the tools. It hasn't widely deployed them yet for soccer.
The strategic implication
For most of the modern era of soccer analytics, the consensus has been that set pieces are noisy and non-repeatable, and that team value is best measured through open-play xG. The Brentford and Brighton results and the now-substantial body of academic work on set-piece-specific modeling suggest that this consensus was always a measurement artifact. Set pieces aren't noise. They're a coherent skill the metrics weren't measuring well. Teams that recognized the measurement gap and invested in the skill outperformed.
The arbitrage on set pieces is narrowing now that more clubs have noticed. Several Premier League sides have added dedicated specialists in the last two years; the traditional French and German champions have done the same. The Brentford-style edge is becoming the league norm, which is what happens to every analytical advantage eventually. The next edge is somewhere else, probably in the second-ball patterns no model is currently capturing and the in-game adjustments that public xG doesn't see. Whoever finds it first will, for a few years, look like a team punching above its weight.