Thursday, 9 April 2015

Deflecting the Blame.

Despite making just two appearances for Stoke City, both from the substitutes bench, Souleymane Oulare's contribution to the Potters' eventual return to the top flight was hugely significant.

His second and final appearance came twenty minutes from the end of normal time in the second leg of the Division Two playoff semi final against Cardiff at Ninian Park.

As had become traditional, Stoke were making a ham-fisted attempt to reach the final. A 2-1 home defeat, combined with a goalless second leg as the game entered the final minute, meant that elimination beckoned.

Cardiff's PA had just urged their fans to remain in the stands to allow a lap of honour from their soon to be victorious team, when Stoke, quite naturally, scored to take the game into extra time.

Five minutes from penalty kicks, Oulare claimed the winner in bizarre circumstances. James O'Connor's free kick was arcing gently into Neil Alexander's left hand corner and the waiting arms of the keeper, when the ball struck Oulare's backside and rolled gently into the opposite side of the net.

Stoke were into the final, ironically to be played at hospitable surroundings of Cardiff's Millennium Stadium, where even the 12 game losing curse of the South dressing room couldn't prevent them from defeating Steve Coppell's Brentford.

Had x, y co ordinates been readily available in 2002, Alexander would have been "guilty" of conceding a speculative 1 in 25 attempt, whereas the reality of the deflection made the shot almost impossible to second guess and save.

Typically around 2% of on target shots take some type of deflection and although the experience of Alexander was extreme, the frequency at which keepers are seen to flail desperately as shots loop unexpectedly from their intended flight before settling in the net, implies that such shots are more difficult to save.

Using data from the recent Premier League seasons, it appears that a third of deflected efforts defeat the keeper's best efforts to save the attempt. Therefore, whether or not a shot was deflected would appear to be a significant additional factor, alongside more usually recorded x,y coordinates and shot type, in determining the likelihood that an effort will result in a score.

A baseline figure for a shot from outside the box increases the likelihood of success from around 0.05 to over 1 in three depending upon the severity of the deflection.

Therefore, a keeper who finds himself facing a disproportionately large number of deflected attempts, may appear to greatly under-perform against a basic shot model that does not include deflections, and their impact on goal expectation for such diverted efforts.

Villa's Brad Guzan faced at least 10 such shots in 2012/13, although the following season was kinder to him. Using combined data from both seasons, a shot model that omits deflections as a predictor, suggested that a single goal was the goal expectation for all deflected attempts faced by Guzan.

Brad Guzan needed all the help he could get to deal with deflected shots in 2012/13.
However, if we re-model all shots to include to added difficulty of saving a deflected effort, Guzan would now expect to concede just over four goals from deflected attempts.

In reality Guzan conceded seven goals from shots that were deflected. An under-performance in both cases, but the keeper's actual concession rate is less damning when we acknowledge the possible effect of a stray deflection may have had on his chances of making what may have been a much easier task.

In a similar vein, Jussi Jaaskelainen over the same period, faced deflected shots that were "worth" 1.7 expected goals from a model that excluded deflections as a variable, but rose to just over 6 expected goals when deflections were accounted for.

He conceded five times from such shots, greatly under performing against a model that fails to differentiate between deflected and non deflected shots, but slightly over performing against a more detailed one that does account for the increased difficulty of saving a deflection.

Repeatability is the key to use data to evaluate players, but a keeper to whom fate hands a glut of deflections to deal with also deserves a model that attempts to allow for the increased difficulty of the tasked he was faced with.

Tuesday, 3 March 2015

Brendan Rodgers' Post January Sorcery?

It's been a "good" week for stats driven narratives. Chelsea's lack of spot kicks, quickly followed by Depay's poor shooting and now Liverpool's post January spurt have all been inferred from mere accumulated data.

The latest story revolved around Liverpool's impressive points per game rate, post New Year compared to their form in the opening five months of the campaign under Brendan Rodgers' tenure.

The figures are neatly summed up below from the excellent source of Liverpool data, Andrew Beasley and although SkySports, to their credit added a ? to their headline describing the trend, twitter was immediately rife with theories behind Rodgers' inspired form in the second half of the campaign.

So are we again most probably looking at definitive conclusions being drawn from raw data, with little or no regard for the effects of random variation?

Only the initial two seasons are complete and although the games are split fairly evenly in the first two campaigns, that does not mean that the early and late games are of similar difficulty.This is most evident in the 2013/14 season.

Liverpool, of course finished runners up last season and in the first half of the season they traveled to play the teams that would prove to be their five closest rivals. Chelsea, Manchester City, Arsenal, Spurs and Everton.

A group of five difficult away fixtures followed, post January by the relatively easier reverse fixtures is likely to have made 2013/14's Aug to Dec fixtures more difficult than the following Jan to May contests.

If we take the implied probabilities from the bookmaker's odds, Liverpool were expected to average around 1.87 ppg pre January and 2.11 post 2013. Even allowing for an inflation of the Reds' rating as the season progressed, it is still likely that the second half of 2013/14 was easier than the first.

But before we attempt to provide reasons for this apparent split performance, (and the various absences of Suarez and Sturridge in early 2013/14 and again in 2014/15 add to the complex mix of potential interactions), we should first see how likely it is that the split appeared just by chance.

If we simulate the two seasons worth of split data and the currently unequal portions from 2014/15, around 1 in 5 trials result in three consecutive seasons where January onwards has been more prolific for Liverpool in terms of points per game compared to the previous five months.

Two seasons out of three where January marks an upturn in points is the most likely scenario, but none of the four possible combinations should be considered as highly unlikely.

This does not mean that Rodgers has not perfected the art of crafting his team into a more effective unit as winter turns to spring or he was unlucky with injuries, but the possibility that we are seeing a split of results that were highly likely to occur for someone, if not Liverpool over the last two and three quarter seasons, simply by chance, should be regarded as a likely contributing cause.