Ex­pected goals ex­plained

FFT dis­sects the anal­y­sis tool

FourFourTwo - - CONTENTS - Words James Maw

Bay­ern Mu­nich prob­a­bly had good rea­son to rue their luck af­ter bow­ing out in the semi-fi­nals of the 2015-16 Cham­pi­ons League against Atletico Madrid – they had lost by the finest of mar­gins. Pep Guardi­ola’s side, hav­ing been beaten 1-0 in the first leg, knew they had to win the re­turn clash in Bavaria by two clear goals. The hosts un­leashed an almighty siege on the Ro­ji­blan­cos’ goal – 33 shots to Atleti’s seven, 11 of which hit the tar­get to the vis­i­tors’ four. Yet, most tellingly, they scored two goals to their op­po­nents’ one and were un­cer­e­mo­ni­ously dumped out on the away goals rule.

The above sta­tis­tics alone hinted Diego Sime­one’s men might have been a touch for­tu­nate, but a more qual­i­ta­tive mea­sure sug­gested that their pro­gres­sion ac­tu­ally bor­dered on the mirac­u­lous.

The next day, speak­ing on Amer­i­can sports net­work ESPN, Ital­ian jour­nal­ist Gabriele Mar­cotti men­tioned in pass­ing that, on an­other night, the Bun­desliga gi­ants would have achieved the re­sult they re­quired to reach the fi­nal – af­ter all, their ex­pected goals rat­ing for the two-legged tie was 4.2 to Atleti’s 1.7.

“You are talk­ing to me about ex­pected goals in the Cham­pi­ons League semi-fi­nal they’ve just lost? What an ab­so­lute load of non­sense,” came the in­cred­u­lous re­ply from pun­dit Craig Bur­ley, the for­mer Chelsea and Scot­land mid­fielder clearly not too im­pressed with the writer’s use of the in­creas­ingly pop­u­lar an­a­lyt­i­cal tool.

“I ex­pect things at Christ­mas from Santa Claus, but they don’t come, right? What I deal in is facts!”

Be­fore Mar­cotti could calmly ex­pand on the finer points of ex­pected goals – or xg, as it’s also known – the ag­i­tated Scots­man let rip again.

“Look at the re­sults! That’s what the game is all about. Whether [or not] you or I or any­body likes it, the game is about re­sults. That is why man­agers get the sack – not all this non­sense about ex­pected goals!”

As video of the heated ex­change went vi­ral, Bur­ley posted on Twit­ter: “Seems I’ve up­set the nerds.”

Mar­cotti, and foot­ball’s bur­geon­ing an­a­lyt­ics com­mu­nity then took a deep breath – fun­nily enough, this was ex­actly the kind of re­ac­tion they had now come to ex­pect.

To the unini­ti­ated, ex­pected goals can ap­pear like lit­tle more than an over­whelm­ingly com­plex equa­tion. How­ever, when you break it down, the very essence of the idea is one fans, pun­dits and man­agers have been sidestep­ping for decades.

“The rea­son I like ex­pected goals is that it’s quite in­tu­itive when you try to strip the math out,” says the writer and an­a­lyt­ics ex­pert Michael Ca­ley, who has been ex­plor­ing ex­pected goals for a num­ber of years. He has shared his dis­cov­er­ies in writ­ten ar­ti­cles and so­cial me­dia posts that have helped pop­u­larise xg among num­ber-crunch­ing sup­port­ers and jour­nal­ists. “Ba­si­cally, it’s the idea of try­ing to eval­u­ate the qual­ity of scor­ing chances,” he ex­plains. “When a pun­dit on tele­vi­sion claims a team was a bit un­lucky and that they could have won a game, what they’re try­ing to say is that the team cre­ated bet­ter scor­ing chances, but the goals just didn’t come.”

It may have only started ap­pear­ing on Match of the Day this sea­son (more on that later), but xg has been around for more than five years and con­tin­ues to be con­stantly re­fined as more matches are played.

“Opta first came up with the con­cept of ex­pected goals when one of our data sci­en­tists – Sam Green, who has since gone on to work at a Pre­mier League club – de­vised an an­a­lyt­i­cal model based on sim­i­lar things be­ing done in Amer­i­can sport,” says Dun­can Alexan­der, Opta’s head of data ed­i­to­rial. “Once the the­ory ex­isted, var­i­ous peo­ple in the an­a­lyt­ics com­mu­nity worked on and ad­justed it – mak­ing a few lit­tle tweaks to the model to try to per­fect it. So there are ac­tu­ally sev­eral dif­fer­ent xg mod­els in ex­is­tence, but there is only re­ally a very slight dif­fer­ence with the num­bers.”

Among those to have tweaked the xg model is Ca­ley, who orig­i­nally be­gan toy­ing with foot­ball an­a­lyt­ics in his spare time while study­ing for a PHD in the His­tory of Re­li­gion at Har­vard Univer­sity. He’s there­fore well placed to ex­plain, in lay­man’s terms, how the whole thing works.

“Ex­pected goals uses a whole bunch of in­di­ca­tors based on Opta’s on-ball event data – where on the pitch the shot had been taken from, what part of the body was used, the type of pass that had set up the chance, how quickly the move pro­gressed down the pitch be­fore the shot, the prox­im­ity of the op­po­si­tion play­ers, and so on – to de­ter­mine ex­actly how likely it is that a par­tic­u­lar op­por­tu­nity will re­sult in a goal.

“For ex­am­ple, if it’s a cross onto a player’s head, that’s go­ing to have lower ex­pected goals be­cause those are more dif­fi­cult to score from. If it’s a through-ball to feet, which is go­ing to elim­i­nate a num­ber of de­fend­ers, that’s go­ing to in­crease the chances of a goal. And if it is a cor­ner-kick, there’ll be a load of de­fend­ers in the box so you’re less likely to score. You es­sen­tially pull all of that into one math equa­tion that then spits out a num­ber – ex­pected goals – which can be tal­lied up over the course of a game, or a sea­son and for a player or a team.”

Crys­tal Palace’s xg for their 1-0 de­feat at Burn­ley in Septem­ber, that ul­ti­mately cost Frank de Boer his briefly held job, was 1.74. Over the course of the 90 min­utes, they spurned sev­eral pre­sentable chances that on an­other day they would have buried. Burn­ley’s xg in the same match was a mere 0.43. The Clarets were ev­i­dently far more clin­i­cal.

At this stage, it is also worth mak­ing a key dis­tinc­tion – that be­tween sta­tis­tics and an­a­lyt­ics. “The thing that re­ally irks me when I hear it is the word ‘stats’,” says Billy Beane – a man who cer­tainly speaks with au­thor­ity. Beane, as many read­ers will be aware, was at the heart of the data rev­o­lu­tion in base­ball dur­ing his time as gen­eral man­ager of the Oak­land A’s. His use of saber­met­rics (“the use of ob­jec­tive data – what we would now call an­a­lyt­ics – and math­e­mat­i­cally find­ing a more ef­fi­cient way of put­ting to­gether a base­ball team”) al­lowed the A’s to go toe-to-toe with Ma­jor League Base­ball’s rich­est fran­chises, de­spite their own fi­nan­cial lim­i­ta­tions. His tale was told in the book Money­ball and the 2011 movie of the same name. He’s also a huge foot­ball fan.

“Stats are re­sults,” Beane tells FFT. “You can have the same out­come, such as a goal, from two dif­fer­ent events but both of them can be very dif­fer­ent in terms of how dif­fi­cult they were. Take a [Lionel] Messi goal, where he has weaved through nine guys, ver­sus a tap-in. Those goals are the same sta­tis­ti­cally, but they re­quire two dif­fer­ent skill sets – one was harder to score than the other.”

Ex­pected goals may now be start­ing to ap­pear in more post-match anal­y­sis along­side shots on tar­get and the num­ber of cor­ners, but it doesn’t re­ally be­long in the same com­pany. While sta­tis­tics will tell you what has just hap­pened, an­a­lyt­ics is able to give you a much clearer idea of what could be yet to come.

“A good ex­am­ple I cite is Ju­ven­tus in 2015-16,” ex­plains Alexan­der. “Af­ter 10 league matches they had only won three times, but over the 10 games they had scored far fewer goals than you’d ex­pect them to have done based on the qual­ity of their chances, and con­ceded more based on the qual­ity of chances their op­po­nents were cre­at­ing. Their re­sults had been much worse than their per­for­mances had sug­gested.

“The Turin side had scored 11 goals in those 10 games, when their xg was 19. At the other end, they had leaked nine, when ex­pected goals sug­gested it would usu­ally have been five. Look­ing at those num­bers, we ex­pected things to regress to nor­mal and, lo and be­hold, the Old Lady’s luck changed. In fact, they won their next 15 Serie A matches on the way to win­ning an­other ti­tle.”

The same method can be im­ple­mented to xg fig­ures for an in­di­vid­ual player. For ex­am­ple, a largely over­looked cen­tre-for­ward who has not found the net too of­ten may be about to start scor­ing for fun – and xg could help you see it com­ing.

“Harry Kane has con­sis­tently scored above his xg for the last three sea­sons,” says Alexan­der. “You are never go­ing to sign a young striker on the ba­sis of one sea­son of sim­i­lar num­bers to Harry Kane, although th­ese num­bers will help you to spot play­ers who, for what­ever rea­son – be it some poor team-mates or a par­tic­u­larly rot­ten spell of luck – may be go­ing un­der the radar.”

Iron­i­cally, Ca­ley – a Spurs fan – was able to use the model to pre­dict Kane’s rise to goalscor­ing great­ness be­fore he had even achieved the sta­tus of ‘one-sea­son won­der’.

“I wrote an ar­ti­cle about Kane’s shot pro­duc­tion be­fore he’d earned a reg­u­lar place in the Tot­ten­ham line-up,” Ca­ley tells FFT. “It out­lined that, in the lim­ited min­utes he was get­ting for Spurs, as well as while out on loan, he had been put­ting up the type of num­bers that looked like those of an elite for­ward.”

Kane’s num­bers dur­ing the fi­nal months of the 2013-14 cam­paign – when Tim Sher­wood was still in charge at White Hart Lane – were, as Ca­ley says, “through the roof”.

It’s not in­con­ceiv­able that, had a shrewd Pre­mier League ri­val taken note of the sta­tis­tics, been a lit­tle bit bolder and made an of­fer for the Tot­ten­ham rookie in the sum­mer of 2014, when he was still very much on the fringes in N17, per­haps he would have re­cently net­ted his 100th goal in their colours in­stead.

But English foot­ball hasn’t al­ways wel­comed change with open arms. Just as for­eign man­agers of the ’90s were met with some be­wil­dered gawps when they dared sug­gest down­ing pints and gorg­ing on steak and chips may not be the per­fect prepa­ra­tion for elite-level ath­letes, those who have more re­cently at­tempted to utilise an­a­lyt­i­cal mod­els to eval­u­ate the game have been met with, at best, a mixed re­sponse.

Poor old Gab Mar­cotti cer­tainly isn’t the first per­son to cite an­a­lyt­i­cal data in assess­ing a sport­ing fix­ture, only to then be im­me­di­ately shot down by scep­ti­cal naysay­ers.

“We were not in­ter­ested in con­vinc­ing peo­ple – frankly it was to our ad­van­tage that no one was con­vinced,” Beane ad­mits to FFT, speak­ing of his early work in base­ball.

De­spite it be­com­ing in­creas­ingly clear that an­a­lyt­ics has got plenty to of­fer, there are still doubters. When xg was made a part of Match of the Day’s graph­ics from the start of this Pre­mier League cam­paign, sud­denly it was main­stream.

Within min­utes of its first ap­pear­ance on screen, so­cial me­dia was in­stantly awash with men­tions of ‘hip­sters and stat nerds’, de­mands for the BBC to ‘get in the sea’ and end­less as­ser­tions that the num­bers are ‘point­less’ and ‘bol­locks’.

This was pre­cisely why, as Match of the Day’s ed­i­tor Richard Hughes ex­plains, the pro­gramme al­ways planned for the in­clu­sion of ex­pected goals to not be too in­tru­sive.

“Match of the Day at­tracts a lot of de­bate on Twit­ter and some­thing new like ex­pected goals will al­ways di­vide opin­ion – that is why we’ve de­lib­er­ately made it a pretty low-key in­tro­duc­tion,” Hughes tells FFT.

“It is there for peo­ple who know about xg al­ready and are keen to see it, but it’s not de­tract­ing from the ex­pe­ri­ence of those who don’t.

“We’ve worked very closely with Opta over the past few sea­sons to in­te­grate a lot more data into the show, and this seemed like a nat­u­ral pro­gres­sion – some­thing new and in­no­va­tive. We have had more and more data on screen – not nec­es­sar­ily things that have been spo­ken about by the pun­dits, but rather sup­port the vi­su­als that have backed up the points they are mak­ing.”

Opta’s Alexan­der con­curs that an­a­lyt­i­cal mod­els such as xg won’t ever re­place liv­ing, breath­ing scouts or pun­dits, but merely aid them.

“We’ve never been zealots,” he says. “We’ve never de­manded that peo­ple use our data or claimed this stuff is go­ing to re­place hu­mans. Ex­pected goals is go­ing to help foot­ball clubs make de­ci­sions and help pun­dits il­lus­trate their point. It’s not go­ing to re­place the hu­man eye.

“Ul­ti­mately, what all th­ese mod­els should do is throw up a lit­tle bit of in­sight and then help peo­ple to form co­gent ar­gu­ments,” he adds.

“I would be ly­ing if I said the pun­dits weren’t a tad scep­ti­cal in terms of the value it brings,” ad­mits Hughes. “Gary Lineker, Ian Wright and Alan Shearer know quite a lot about scor­ing goals, and there have been vari­ables in the model that they’ve ques­tioned when we’ve dis­cussed it – in par­tic­u­lar, things such as de­fen­sive po­si­tion­ing and long-shot chances. The key for them is al­ways which player has taken the shot.”

So the strik­ers’ union will al­ways have their say on the per­for­mances of their brethren – re­gard­less of the rise of xg – but what about other ar­eas of the pitch? Will we end up hav­ing some sim­i­lar con­ver­sa­tions about de­fen­sive con­tri­bu­tions?

“Events on the ball are what we all fo­cus on, but there are so many other things go­ing on that will im­pact what hap­pens next,” ex­plains Beane. “There are things that hap­pen on a foot­ball field that aren’t be­ing mea­sured, so play­ers don’t get the credit for them. For ex­am­ple, a de­fen­sive player, who by virtue of his abil­ity is able to get him­self into a po­si­tion to al­ter a shot, will com­pletely change the dy­namic of the play de­spite never touch­ing the ball. Even­tu­ally, that is the kind of thing you want to mea­sure.”

The good news for Beane and the world’s best cen­tre-backs is that an an­a­lyt­i­cal way of assess­ing de­fen­sive con­tri­bu­tion is in the pipeline.

“Ex­pected goals is the first model and the one that has re­ceived the most cov­er­age, but it’s the first in a series of hope­fully quite a few we will be us­ing,” says Alexan­der.

“We’re also now work­ing with ‘ex­pected as­sists’, which is sim­i­lar to xg, and ‘se­quences’ from which you de­rive a team’s style of play and the pace at which they at­tack.

“And we’re also work­ing on some­thing called ‘de­fen­sive cov­er­age’, which could be big for us be­cause the crit­i­cism of Opta event data – and a rea­son­ably valid one – has been that it’s a lot harder to as­sess de­fend­ing than it is at­tack­ing.”

De­fen­sive cov­er­age can mea­sure the area of de­fen­sive re­spon­si­bil­ity im­plied by a player’s de­fen­sive ac­tions through­out a match – tack­les, blocks, in­ter­cep­tions, clear­ances etc. So Chelsea’s all-ac­tion mid­field lu­natic N’golo Kanté, for in­stance, may cover a large area of the pitch, while a full-back in a team that’s be­ing dom­i­nated by the op­po­si­tion will likely have a smaller area.

“A good ex­am­ple of that from last sea­son was when An­der Her­rera marked Eden Hazard out of the game [be­tween Manch­ester United and Chelsea] at Old Traf­ford in April,” says Alexan­der. “He’s nom­i­nally a cen­tral mid­fielder, but the Spa­niard’s ‘de­fen­sive zone’ was a rough par­al­lel­o­gram on the edge of the right-hand side of United’s box. He was tasked with stop­ping Hazard, who ul­ti­mately didn’t have a sin­gle touch in­side the penalty area.

“Any pun­dit who watched that match would cer­tainly have spot­ted that Her­rera per­formed very well, but up un­til now there hasn’t re­ally been a way of il­lus­trat­ing that.”

That may not be mu­sic to the ears of Craig Bur­ley, half of Twit­ter and any­one else who’d rather stick their fin­gers in their ears and pre­tend foot­ball’s ‘data rev­o­lu­tion’ isn’t ac­tu­ally hap­pen­ing. But as Billy Beane puts it, “the ge­nie’s out of the bot­tle now, and it’s not go­ing back in”.

CRYS­TAL PALACE’S XG In THE 1-0 DE­FEAT AT BURN­LEY THAT COST FRANK DE BOER HIS JOB WAS 1.74, WHILE THE CLARETS’ XG In THE SAME GAME WAS A MERE 0.43

Clock­wise from top The Ea­gles were un­lucky at Burn­ley, ac­cord­ing to

xg fig­ures; the an­a­lysts pre­dicted Juve’s re­vival

dur­ing 2015-16; Beane, who led the rev­o­lu­tion in base­ball an­a­lyt­ics, feels a sim­i­lar ap­proach could help foot­ball scouts, too

Newspapers in English

Newspapers from UK

© PressReader. All rights reserved.