Manchester City was in trouble. The team had dominated the Premier League by passing its way into perfect shooting positions, but if it couldn’t find a way to score against Leicester City in the penultimate game of the 2018-19 season, Liverpool could snatch the title away. In the 70th minute, center back Vincent Kompany trundled the ball into the attacking third. Instead of laying it off to a teammate, he saw an opening, put his head down and let rip from 30 yards — an embarrassingly low-percentage shot.
It was the kind of chance modern players don’t take as often as they used to. Expected goals (xG) models show that the probability of scoring increases sharply as the shooter gets close to goal, and as xG has gotten popular, shot distances have fallen in league after league. City manager Pep Guardiola later admitted that as Kompany cocked his leg back, he was thinking, “Don’t shoot! Pass the ball!”
Van Roy et al.
But is shooting from distance always the wrong choice? The decline of shots from outside the box is often described as an analytics-driven shift, like the death of the midrange jumper in the NBA. In basketball, however, a team that passes up one shot will usually find another on the same possession. Shot selection in soccer is a murkier problem. “The potential payoff of not shooting is that an even better shot may arise down the line,” explains a new paper at this week’s MIT Sloan Sports Analytics Conference, “but there is no guarantee of this happening.”
The paper’s lead author is Maaike Van Roy, a Ph.D. student at the Belgian university KU Leuven, where a team led by professor Jesse Davis does soccer research focused on artificial intelligence. “Outside the penalty box, the question is, ‘Should you shoot immediately, or should you move?’” Van Roy said. To answer that question, Van Roy and her colleagues devised a way to compare the xG of a potential shot from distance with the probability of scoring one or two moves later after passing up the shot.
“We found specific long-distance zones where teams should be shooting from instead of moving,” Van Roy said. “We also found that when you increase the number of movement actions [after declining to shoot], you’re less likely to generate a goal, especially when you’re passing to the flanks.”
In other words, even a shot that’s expected to score just a few times out of a hundred may not be such a bad idea if every pass, carry and cross to try to work the ball into a higher-value spot represents another coin-flip chance of losing the ball. It’s possible Kompany’s gamble from outside the box against Leicester was a better bet than Guardiola thought.
To estimate how likely a team would be to score if it tried different things, the researchers used the Markov decision process to divide the attacking half of the pitch into a grid of tiny zones and calculate how moving the ball from one zone to another would change goal probabilities. They trained 17 different models, one for each English club that appeared in the data for both the 2017-18 and 2018-19 seasons. That turned up some interesting differences in optimal behaviors.
Related: Club Soccer Predictions Read more. »
“For some teams, you have these zones to the left of the penalty arc” where it’s better to shoot than to attempt a pass or carry first, Van Roy explained. “We could actually see that those teams that have these zones also have left wingers that are right-footed. We think that a reasonable explanation is that they could cut infield, get to their good foot, and then shoot from those locations.”
“Like Eden Hazard with Chelsea,” Davis added.
Whether or not they played with inverted wingers, every club had some central areas at the top of the box where shooting offered a better scoring chance than trying to get closer. The researchers investigated what would happen if a team changed its long-shot policy, for example by shooting 20 percent more often from the zones where the model preferred it, and they estimated that it could add about 1 extra point over the course of a season. That may not sound like much, but it’s the margin that won Man City the 2018-19 Premier League title after Kompany’s shot flew up and into the top corner of Leicester City’s net.
There’s still a lot to sort out before anyone tells Guardiola to overhaul his attacking tactics. The Sloan paper treats any interruption in the attacking move, even a shot rebound or a corner kick, as the end of a scoring chance. A team that passes into the box may not be counting on a clean shot right away and would happily settle for a goal off a loose ball or a bad clearance. The more messy contingencies the game throws at researchers, the harder it gets to measure trade-offs precisely. In that respect, the paper’s focus on scoring possibilities no more than one or two moves ahead mimics what a shooter can foresee. “Two or three actions out is probably what most players are able to think about anyway,” Davis said, “and after that it’s probably a lot of guesswork.”
Closer shots are a good thing if you can get them, sure. That’s the basic lesson of expected goals. Comparing the risk and reward of in-game attacking options is harder, but it could be more useful — and more fun. Davis and Van Roy’s team built an interactive tool to let people get a feel for some of the numbers around long shots. “We want to have people ask questions to the models, like what would happen in this situation?” Davis said.
The hope is that better information will help players make the right choices in the heat of the moment. “The artists are the players. They have to decide in one fraction of a second,” Guardiola told a reporter who asked about Kompany’s shot. “He made a good decision.”
Check out our latest soccer predictions.