How Big A Shock Was Bradford 3 Aston Villa 1 ?

Seventy places may separate Aston Villa from their deserved League Cup conquerors Bradford City, but few would claim that the Yorkshire side should now be considered superior to Paul Lambert's beleaguered kids. The seventy place gap may not accurately reflect the current gulf in class, but Villa are still superior to Bradford and the second leg may confirm this view.

What Tuesday night's game does demonstrate is that teams can and do produce performances that are someway above or someway below the average level of their usual performance. Football is a low scoring sport and therefore random chance will be a big contributor to a single match result. Over a larger run of matches a side's true levels of talent will begin to come to the fore in the win/ draw and loss column and the best teams will show average levels of performance that more accurately reflects their real ability. That Bradford can defeat Villa merely indicates the levels to which Villa's standards can fall and Bradford's can rise to on one particular gameday in an environment where a couple of outstanding saves can be immediately followed by a second goal for the less accomplished team.

To visualize the peaks and troughs that can be expected in leagues and for particular teams it is first necessary to establish an expected level of performance for each team based on reliable indicators, ideally recorded over a reasonable length of time. There are numerous ways to forecast the most likely outcome when two sides meet at a particular venue ranging from methods involving goal scoring and conceding rates corrected for opponent strength to correlating performance indicators with previous match outcomes. Invariably, and most usefully we can eventually express the likely outcome of a match in terms of the average number of goals one team would be superior to their opponents should the game be repeatedly replayed.

Any reasonable proficient prediction system should eventually mirror a side's actual performance, especially if a continual under or over performance is taken as a sign that the overall strength of a team has shifted. If we are satisfied that we have a good prediction method, we can use a side's match by match goal difference compared to the pre game, expected goal difference to show the extremes of performance, both good and bad that typical teams are capable of.

For example if a team is a 1 goal favourite and wins by three clear goals, they could be considered to have over performed by a margin of two goals, while a draw would indicate under performance to the tune of one goal. By collecting and plotting the frequency of such outcomes, the spread of a team's actual performance can be charted.

Above I've plotted the frequency at which Tottenham beat or failed to beat a goal based pre game estimation of their chances in each match from the 2010-11 EPL season. The residual expectations are collected in half goal bands, so the zero column counts each occasion where the actual margin by which Spurs won or lost the game fell between 0.25 of a goal above or below expectation. The column immediately to the right of the nearly central zero column consists of the six occasions during the season when Spurs over performed by between 0.251 and 0.75 goals when compared to their expected pre game performance.

The majority of Tottenham's performances in 2010/11 cluster around their pre game estimates, a sign that the the model used is tracking reality fairly well. Two matches, a 1-0 defeat at home to Wigan and a 3-1 defeat away at Blackpool fall into the category at the extreme left comprising matches were the Spurs side under achieved by between 2.25 and 2.75 goals.

Overall the model does a good job of describing Tottenham and one season's worth of results produces over 20 results that were predicted with reasonable pre game accuracy. The half dozen games at either extreme shows the frequency with which Spurs turned in atypically good or bad performances and the distance from the zero line indicates roughly how extreme those outcomes actually were.

If we plot a similar graph for every EPL game played by every home team since the beginning of the 2005-06 season, a similarly well defined curve results. The majority of matches fall very close to the average margins of victory predicted beforehand.

 Again the buckets into which each game is placed are half a goal wide and once again the more extreme actual margins of victory or defeat compared to pre game estimates appear with increasing scarcity. Aston Villa were considered a shade over 1.2 goals superior to Bradford on most reliable indicators before kick off at Valley Parade, so a defeat by two goals represented an under achievement of 3.2 goals on the night or a similar over achievement by their hosts. Using the crude sampling bins, that would have placed the Bradford Villa result into a bin comprising home teams who had played between 2.75 and 3.25 goals better than expected, as indicated by the arrow on the plot above. An occurrence which was played out just 26 times out of 2280 matches in the EPL between August 2005 and May 2011.

How Well Do Individual Scoring Rates Survive A Change Of Scenery ?

The opening of the transfer window always brings with it a huge amount of speculation concerning the possible destination of unsettled, out of favour or much sought after players. Many of these apparently "done deals" fail to materialize and the only real certainty surrounding the January spending spree is that the window will "slam shut" at the end of the final day.

Previously much of the pairing of players to their ideal destination was done on gut feeling, but the ever increasing amount of individual data  has now enabled even the casual fan to pick out that tough tackling midfield enforcer who will propel his side to either safety or greater heights. So has the readily available shooting, scoring and tackling statistics made completing the jigsaw an altogether simpler process ?

Goals are widely used to define strikers and in this post on super subs, I used the number of goals scored by Dzeko for Manchester City as a proportion of all goals scored whilst he was on the pitch in an attempt to quantify his contribution in both roles. This method is attractive because it partly accounts for quality of opportunity, a rout against a poor side will see other players also getting on the score sheet and it also allows for the heightened rate of scoring later in matches. Also a player scoring two goals in a first half isn't penalized if he is absent for the second period and further goals are added to the total.

By normalizing the scoring environment, while Dzeko's team mates remain reasonably constant, we can try to judge if there is likely to be any difference between Dzeko, the starter and Dzeko, the sub. We can further develop this approach to try to estimate how a newly acquired striker may fit into a new team, especially as measured by the bottom line of goals scored.

Estimating player ability is always difficult, ageing and survivor bias for example is hardly ever addressed, but the most glaring problem in this instance is that transfer targets are playing for a different team, alongside players of differing abilities and varied tactical priorities compared to the side which is pursuing him. A player may score a high proportion of the goals claimed by a struggling Premiership side, but that may be because he is head and shoulders above his striking team mates. The numbers may be telling you that he is "too good" for his present side, but is he good enough to play for a potential suitor ? A potential buyer needs to know how he is likely to perform for them and raw scoring exploits are unlikely to provide reliable conclusions. Are you buying a squad player or an upgrade ?

The case of Daniel Sturridge highlights how we may be able to make more informed predictions using the  scoring exploits a player achieves whilst playing at different clubs. Sturridge has performed in the top flight at the then mid table Manchester City, Championship winning Chelsea, struggling Bolton, by virtue of the loan system and now appears to have found a level where appearances and the opportunity to showcase his talent will be guaranteed at Liverpool.

During his three seasons at City, he scored 30% of his side's goals whist he was on the pitch. At Chelsea the figure dropped to just below 20% and in a half season loan at Bolton it shot up to nearly 50%. Sturridge's opportunities have been limited, so sample sizes are small, but his "talent" as a striker appears to rise as the overall quality of his side falls. His proportional strike rate was exceptional at Bolton, very good at City and good at Chelsea.

The reality is probably that he was close to being the same player at all three clubs, especially during his time at Chelsea and Bolton. He shone outstandingly at The Reebok because he was the best attacking player on the team and as such was able to dominate his fellow strikers. Once he arrived back at Chelsea he was partly eclipsed by better strikers and his proportional rate of goals scored fell. Had Chelsea purchased, instead of merely recalling Sturridge from Bolton on the basis of his 50% strike rate, they would have been disappointed if they had expected that headline stat to be repeated at The Bridge.

In the table above, I've added the proportion of goals scored by a variety of strikers who have played for multiple seasons at Premiership clubs of differing overall quality. The strikers included Crouch, Adebayor and Bent in addition to Sturridge and the clubs involved range from Chelsea and Arsenal, down to struggling EPL sides like Southampton and Charlton. The relationship which appears to exist for Sturridge also shows up in this larger group of similarly talented strikers, each of whom have been traded for eight figure transfer fees over the course of the sample. They tend to score a smaller proportion of the goals scored as they move to bigger and better clubs.

When playing for struggling EPL sides, denoted by success rates in the region of 0.35, the line of best fit indicates that such strikers are likely to account for between 40 and 50% of the team goals scored whilst they are on the pitch. When making the step up to a Champions League quality side, their contribution then typically falls to around 20% because they are probably surrounded by a bigger pool of goal scoring talent.

Crouch, Scorer of 22% of Liverpool's Goals when playing compared to 35% for Stoke.

Collecting the proportional scoring records of players provides an added level of information which may prove valuable in predicting future performance at other clubs. Knowledge of a player's record at his previous clubs may enable a side looking to buy in the transfer market to estimate the proportion of goals a new signing would typically contribute at a different class of club. And more importantly whether this figure will go hand in hand with an overall increase in total goals for the team. Sturridge's presence in the Bolton lineup also coincided with a 10% increase in total goals scored compare to their previous 30 games, although this needs to be confirmed in much larger sample sizes.

Demba Ba's record at West Ham and Newcastle has seen him score around 40% of his side's goals as an active player at teams with a combined success rate in the region of 0.44, well in line with expectations. If he continues to hug the regression line, Chelsea may have purchased a player capable scoring 15% of the goals at a club such as Chelsea. Further research may show whether this will be an upgrade at the London club, although the reduced price tag for such an apparent talent would already appear to be good value.

Luis Suarez, Daniel Sturridge's intended strike partner at Liverpool also fits neatly onto the line of best fit. He has scored around 30% of the goals when he's played at Anfield, just the figure you would expect for a top striker playing for a team with an overall success rate of 0.54 spread over two part and one completed season. The figures for Sturridge and Suarez therefore point to Liverpool having now acquired two top class strikers. The hope will be that the pair will produced strike rates nearer to 20%, provide scoring opportunities for other team mates through assists or by occupying opposition defences and Liverpool's success rate as measured in wins and draws will increase in line with the profile of their new striking threat.

Liverpool's previous big money swap saw Andy Carroll join from Newcastle. Carroll and Newcastle spent 2009/10 in The Championship, but during his time playing in the Premiership, Newcastle's success rate hovered just below 0.4 and a top striking prospect playing for such a struggling side should have scored around 45% of the team's goal. Carroll only accounted for 35%, so on that basis he was a much bigger gamble than the one they've taken on Sturridge, who cost less and has better numbers.

This methodology is reasonably straightforward when used for goals, but it is equally feasible to apply it to add context to other on field actions that could otherwise potentially lead to misleading conclusions. Looking at the number or even the proportion of successful passes or assists may be just the start of the selection process and not the endgame.

Also check out Danny Pugsley who takes at look at the subject here

The Usefulness Of Tackles.

Anyone who doubts the elevated regard in which attacking play is held in comparison to defensive talent need only watch an episode of Match of the Day. The day's worst defensive performance invariably appears first in the running order, while the best defensive show is aired, usually with poor grace in the show's final minutes.

Analysis of attacking play is also more heavily developed compared to the defensive side of the ball. Goals are the obvious end product of attacking intent, they alone decide the final match outcome and there is a well defined and easily quantifiable chain of evidence leading from goal creation to realisation.

Shots and shots on target are one step removed from an actual goal and correlate well with winning matches.  Assists come next in the chain, again providing a comfortingly strong correlation with success and it is only when we step even further back into the scoring process to look at mere, run of the mill passes that we begin to experience a more confused correlation between quantity and quality of execution and end product.

At the sharp end of goal scoring there is no compromise, a goal is scored (usually) when the ball crosses the line, but how we arrive at the successful conclusion can take on a multitude of different tactical approaches. Therefore, we eventually reach a point in the scoring process where there is no catch all statistic that can describe with equal clarity a possession based Barcelona approach or a direct, long ball plan.

Shots and assists can be used to illustrate a team's previous successes and partly predict their future expectations, but passes are more indicative of team tactics and such diversity often defies easy or relevant measurement.

Defensively the evidence chain is much shorter and potential for confused and weak  correlation is reached much quicker. Interceptions and tackles are widely regarded as the currency by which defenders are valued, but the reality is that these stats are much more the product of how a team is setup than how talented are a side's defenders. Tackles are the passes of the defensive world, they are not the equivalent of shots or assists. Tackles per minute were an irrelevance to the likes of Paolo Maldini because there are more ways to prevent a goal than there are to score one.

Bolton, A Successful Tackle or Just Buying Time?

A good starting point in trying to understand the part tackles play in the game is to see how often teams commit to a tackle. The risk involved in tackling is twofold. A mistimed effort can lead to cards and a failed attempt usually sees the tackler out of position, if not removed entirely from the rest of the attacking move. Sorting tackles by outfield playing position can give reasonable picture of where a team is making their challenges and in the table below I've corrected the figures to allow for playing time. Defenders or midfielders accrue over twice the playing time of strikers, partly through tactical substitutions and because they are more numerous in the lineup.

A straightforward totting up of tackles would therefore always see strikers trailing the other two positions, so I've corrected cumulative totals from the 2011/12 season to produce normalized per game figures for all three positions. The numbers should depend on the amount of tackling demanded of each position by each club and the requirement imposed on them by the opposition.

         Tackling Rates Per 90 minutes by Position, EPL Sides 2011/12.

Team.	Defenders.	Midfielders.	Strikers.
Aston Villa.	24	23	9
Manchester United.	23	25	10
Arsenal.	22	21	9
Norwich.	22	19	6
Manchester City.	22	18	10
Chelsea.	21	22	10
Stoke City.	21	18	9
QPR.	21	23	8
Wigan.	20	21	7
Sunderland.	20	25	10
Swansea.	20	22	9
Bolton.	19	21	16
Newcastle.	19	25	8
Liverpool.	19	22	13
Spurs.	19	23	10
Wolves.	18	17	10
Blackburn.	18	20	13
Everton.	17	19	8
WBA.	16	23	9
Fulham.	15	19	19

The baseline figure for team tackles is just under 19 per game. So Fulham's strikers were asked to tackle back at near league average team rates last term, almost twice the rate required of strikers as a group. A quality they may have sacrificed this term with the acquisition of Berbatov. They were followed in the tackling stakes by Bolton, anecdotally a side which defends from the front. In contrast, Stoke strikers are expected to contribute well below average tackling numbers, although regular Stoke watchers will know that Pulis demands pressure before contact as a tactic. Indeed Stoke's midfield also had the second lowest number of attempted tackles in 2011/12 and are only just in the top half of the rankings for defenders. Tackles are a component of defending, but not the only one and in the case of some teams such as Stoke, they are not the primary one.

Manchester United's commitment to tackles in 2011/12, both overall and particularly from defenders and midfielders is well illustrated. Last year's runners up required their midfield players to be able to tackle often, an attribute used by a variety of other teams ranging from Villa, Sunderland, QPR to Spurs, Chelsea and Newcastle. By normalizing tackling stats to the equivalent of a team consisting entirely of strikers, midfielders or defenders, season long playing styles by position become more transparent and may better highlight qualities required from various sides.

Overall there is virtually no correlation between tackle numbers as frequently exemplified by tackles per minute figures for individual teams and the ability to prevent goals, again indicating that raw tackle numbers are a product of tactics rather than proficiency. We need to look at tackling efficiency to see how effective these tactics are and how successful tackles may contribute towards goal prevention and ultimately help teams to win.

A successful tackle at worst slows down an opponents passing sequence and ideally teams would like to see a positive correlation between efficient tackling rates and goal prevention. Overall tackle rates usually lie between 70% and 80%.

          Percentage Tackling Success Rates by Position, EPL 2011/12.

Team.	Defenders.	Midfielders.	Strikers.
Arsenal.	75.1	72.9	80.8
Aston Villa.	73.9	76.9	79.5
Blackburn.	72.8	75.5	80.0
Bolton.	75.3	69.9	77.1
Chelsea.	77.1	76.5	82.8
Everton.	75.1	80.9	75.0
Fulham.	78.5	76.0	76.9
Liverpool.	74.0	73.2	75.3
Manchester City.	75.7	73.2	67.7
Manchester United.	77.1	71.3	87.3
Newcastle.	75.3	73.0	75.9
Norwich.	68.7	73.7	78.4
QPR.	73.1	72.0	69.2
Stoke.	72.1	68.0	72.0
Sunderland	78.1	75.7	69.6
Swansea.	73.8	71.2	71.0
Tottenham.	78.8	71.6	69.2
WBA.	73.4	77.6	78.8
Wigan.	72.3	73.4	85.3
Wolves.	71.9	76.8	80.6
Overall.	74.6	74.0	76.9

It's initially surprising to see strikers as the most efficient group of tacklers, with Manchester United's attackers particularly out stripping their defensive team mates. However, this highlights the danger of taking statistics at face value. Strikers are out numbered by other positions by around two to one in on field presence, so the smaller sample sizes are more likely to lead to extreme value either above or below the norm. Hence United's 87% success rate may not be repeated in larger samples, bringing rates closer to those enjoyed by the defence.

The higher overall figure for strikers as a group is likely to be because they can pick the tackles they attempt and sometimes opt for other defensive tactics, such as forcing a pass by closing down an opponent. Midfielders and particularly defenders are often required to attempt tackles that are the last ditch alternative to allowing a decisive pass or shot on goal. In short, tackle difficulty isn't well represented in these raw figures.

The correlation between team tackle efficiency sorted by position and goal prevention is non existent for strikers and possibly surprisingly, also for midfielders and it's only when we look at just defenders that a reasonable  correlation develops.

The efficiency of defensive tackles is the strongest tackling indicator of goal prevention beating out overall rates, rates for the other two nominal positions and raw tackle counts for defences. The relative weakness does indicate that other factors, most probably interceptions, applying pressure to induce errors and goalkeeping ability are also significant factors.

We can use a similar approach from an attacking perspective by recording the rates at which opponents complete tackles against individual teams to see if an ability to make defenders miss tackles leads to increased scoring rates.

The correlation in this instance is weaker, but the indication is that an ability of a team to make defenders proportionally miss more tackles tends to bring rewards with increased scoring. Low scoring sides from 2011/12 such as Stoke, Wolves, Villa, Swansea and Sunderland allowed opponents to successfully complete upwards of 75% of tackles attempted by defenders, while more free scoring sides such as Newcastle, Spurs and Arsenal made defenders miss more frequently. Both Manchester clubs were big outliers, accounting for the reduced r^2 values, with City particularly being capable of scoring extremely freely despite facing near average tackle rates from opposing defenders. Possibly an indication that City excel at overcoming other defensive tactics which are omitted from mere tackling data.

Both graphs can be combined to see how important tackles made and tackles received may be to the majority of EPL sides. Teams such as Chelsea, Arsenal, Manchester United, Spurs and Newcastle had defenders who tackled at above average rates and had attackers who induced opposing defenders to tackle at below average rates. This seemingly heady mix was translated into impressive finishing positions. However, Bolton and to a lesser degree QPR produce similar splits, but without the successful results. So once again we could conclude that tackle statistics are noteworthy, but incomplete predictors and possibly partly flawed.

To hint at why tackles aren't as indicative of success as say shots, we can look at Bolton's game against Manchester City from last term. Overall, Bolton made 19 tackles, 16 of which were deemed successful for a tackle rate of 84%. However, in just 11 of those 16 successful tackles were Bolton able to gain possession and try to develop a footballing move of their own. On five occasions Manchester City retained possession because the ball was put out of play for either a corner or a throw. It's churlish to deem an excellent touchline tackle as a failure, but tackles do come with varying degrees of success. Five times City were able to continue the move from very close to where a successful tackle was made, dropping Bolton's success rate from  84% to 58%. Similarly for City, they recorded nine success from ten tackles, but on three occasions Bolton retained possession from the resulting throw and a 90% strike rate fell to 60%.

Just as shot based models improve as we add details such as x, y co ordinates, tackling models may benefit from reference to the immediate aftermath of the challenge.

Defence is always more complex to describe numerically. Ways to defend your goal are more numerous, varied and sometimes in the case of pressing, difficult to record compared to scoring which ultimately boils down to a shot or header at goal. Tackling efficiency appears to be preferable marker compared to frequent but less successful tackle attempts and by grading different levels of success for individual tackles we may creep closer to identifying the great defenders from the merely good. While appreciating that some teams demand different approaches to defence where tackling may not be the primary priority.