![](https://omsstats.wpenginepowered.com/wp-content/uploads/2020/03/GettyImages-465007989-800x569-1-1000x380.jpg)
![](https://omsstats.wpenginepowered.com/wp-content/uploads/2019/05/icon-soccer-ball-16x16.png)
This article is a written version of Garry Gelade’s 2017 OptaPro Analytics Forum presentation on crossing. Garry’s work featured in the Guardian and Garry also conducted interviews on both the Keys and Gray Show and TalkSport.
![](https://omsstats.wpenginepowered.com/wp-content/uploads/2017/08/fig1.png)
Figure 1. 1000 random crosses
What people think and what the numbers prove
A great cross is one of the most memorable sights in football, but is crossing an effective form of attack? Some say it is now out-dated in the modern game and remains a fossilised English obsession. It is true that the past six years have seen a decline in the number of crosses in top European football from around 17.5 per match in 2010-11 to 15 in 2016-17; but the numbers show no difference between the Premier League and other top European leagues.
Some claim that crossing is ineffective, arguing that it takes 92 crosses to score a goal. But that number only considers the immediate effect of a cross. When we acknowledge the indirect effects of a cross (goals scored from the second ball, corners and penalties conceded) we find that it needs only about 45 crosses to produce a goal – a much better return.
The most extreme critique of crossing has come from the Czech economist Professor Jan Vecer. Vecer argued that crossing is not only ineffective but counter-productive. In a detailed statistical analysis he showed that teams that cross less score more goals, and he concluded that if Premier League teams stopped crossing altogether, they would score an extra 15 goals per season.
Unfortunately this argument embodies a causal misconception, as we can see when we look at how the scoreline evolves over the 90 minutes. I analysed changes in crossing rates using the framework shown in Figure 2. I divided the games into segments at the points where goals were scored and examined how the frequency of crosses changed according to whether a team scored or conceded a goal.
Figure 2. Analysis of changes in crossing rate
The results were clear: crossing decreases after a team scores a goal, and increases after a team concedes a goal.
Teams don’t score more goals because they cross less – they cross less because they score more. And that is quite a different thing. It makes footballing sense; a team in the lead will usually play more defensively than when they are behind. There is no evidence that reducing crosses will increase the number of goals scored.
Finding the most effective type of cross
What is the most effective type of cross? Where should a team cross from, and where should they direct their crosses? I analysed 33,954 crosses from three seasons of the Premier League, of which 666 were successful (a cross was deemed successful if it was followed by a goal within the next six seconds).
To classify the crosses I used a statistical technique called the Conditional Inference Tree. To show how this works, I applied it to an example slightly removed from crossing in professional football. I’ve applied it to finding categories of survivors from the Titanic disaster, where the great ‘unsinkable’ ship with 2120 souls on board sank in the icy waters of the Atlantic in February 1912. The tree algorithm discovered nine categories of survivors as shown in the diagram below.
We can see for example that survival rates of women tend to be higher than those of men. The survival rate of women travelling first class (.97) was about twice as high as that of women travelling 3rd class. Only 16% of men travelling 3rd class and 34% of men travelling 1st class survived. Truly a case of women and children first.
The conditional inference tree is particularly useful when the number of successes in the data sample is small compared to the number of failures (in fact the more traditional type of regression tree which I tried first, failed to detect any signals at all).
Applied to our crossing data, the Conditional Inference Tree discovered distinct categories of cross which varied according to their start and end locations, whether they were chipped or driven, and which had success rates ranging from 0.2% to 11.8%. The crosses with the lowest success rates were those that were either too short or too long, in other words passes landing beyond the far post or short of the near post. The most successful category of cross was a driven pass that just crossed the centre line by no more than five metres.
The images below illustrate some of the other cross categories, and highlight the difference in success rates that a small change can make.
Figure 3. Categories 6 and 7 compared.
Categories 6 and 7 consist of crosses into the box landing in a region about 4m either side of the near post and between the 18 and six yard lines. Where they differ is the origin of the cross. Category 7 crosses are delivered from deep – behind the penalty box – and have a conversion rate of only 0.5%. Category 6 crosses are delivered from further up the field, which more than triples the success rate to 1.7%.
Figure 4. Categories 17 and 18 compared
Categories 17 and 18 consist of crosses chipped to the far post originating at least five metres from the touch line and end up no more than five metres from the centre line. The difference between the two is distance from the goal line of the end location. Crosses in Category 17 land behind the penalty spot, while crosses in Category 18 end up between the penalty spot and the goal. This seemingly quite small difference has a big influence on the success rate. The conversion rate for crosses delivered behind the penalty spot is only 2%, but crosses delivered in front of the penalty spot have a success rate of 5.8%.
Application
Together with a representative from a Premier League club, I divided the area from which crosses are most often delivered into four zones (shown in Figure 5). I then consulted the data to make recommendations for maximising success rates from each zone.
Figure 5. Four crossing zones
Figure 6a. Recommendations for crossing from zone 1
Figure 6b. Recommendations for crossing from zone 2
Figure 6c. Recommendations for crossing from Zone 3
Figure 6d. Recommendations for crossing from Zone 4
Conclusions
Crossing got something of a bad name amongst observers of the English game after the failure Damien Comolli’s attempt to build a team based on crossing at Liverpool. With great crossers of the ball like Henderson, Enrique and Downing along with Andy Carroll waiting in the box, Liverpool produced 787 crosses in the 2011-12 season – far more than any other team – for a meagre total of four assisted goals.
As I’ve written here, the problem was likely that the overemphasis on crossing led to crosses being delivered too early; Liverpool’s crosses originated about three metres further up the pitch than other Premier League teams.
Then again, Liverpool’s style might just have been too predictable. It’s much easier to defend if you know how your opponent is going to play, and common sense suggests it’s good to have a variety of attacking options. If you never cross the ball, or never play through the centre, your opponents have a better chance of nullifying you. So although crossing in isolation may not be the most effective way to score goals, I think it’s important to keep it in the mix.
Most people I spoke to – both during and after the OptaPro Forum – were relieved to find that crossing still has a role in the modern game. The feedback from coaches and clubs was also very positive, and they appreciated the demonstrable link between the analytical findings and actions on the pitch.