Passing failures are as important as successes

Passing – failures are as important as successes

I’ve been thinking a lot about passing recently.

It seems to be the area where I think there is most room for improving what analytics can offer.

The traditional response to passing percentage statistics is “anyone can pass the ball under no pressure to an unmarked teammate”. That is true when you reach the professional level. It is also true that the players with the highest pass percentage success rate are generally centre backs and deep midfielders. They generally play short passes, in less crowded areas to unmarked players.

So what is the state of play now?

The standard player level data collection output contains:
Number of passes
Success rate of passes
Passes forwards, sideways, backwards
Success rate of each type of pass
Through balls (not sure on definition)
Key passes
Passes into box (open play and all)
Long balls
Final third passes (forwards, backwards, sideways)

This is all great to have but you can start to pick holes in almost all the definitions and come up with counterexamples straight away. For example, what is a long ball? A Beckham 70 yard switch into space is different from an aimless hoof. You want a player with lots of forwards passes in the final third? OK you’ve lost Sterling and Sane then who play lots of backward passes, these are also called cutbacks and are one of the best ways to score.

The biggest problem with this aggregate data is that it still doesn’t really tell you if you are making the right pass. We might want forward passes but a forward pass to a player under loads of pressure, even when complete, might be worse than one to a teammate in space to your side.

For that we probably need to know more details about where you are on the pitch when you are making these passes and where they are going to.

Paul Riley has produced a good passing model which I covered recently. That looks at pass success from X/Y coordinate to X/Y coordinate. This is a good step forward as you can the types of passes a player is making, whether they are above or below average difficulty and if the player is making them at above or below average completion rate.

I really like the simplicity of the model, it is a big step forward and passes the “smell” test.

What it doesn’t adjust for, which I think is important, is the relative dominance and skill of the player’s team. Take one player from a poor team that plays directly and put them on a possession-based team of superstars and their passing statistics will change dramatically. Opponents play differently against Manchester City to Cardiff. Yes, John Stones is a much better passer than Sol Bamba but he also has much better options to pass to, a manager who wants him to pass out from the back and probably more space to play into as opponents hunker down for a low block.

A good step forward for this is Statsbomb’s pressure index which includes the ability to look at how many passes players play under pressure and changes in their passing style when under pressure. Again the results from this are all up for interpretation, for example, I would expect a dominant possession based team would play more passes under pressure higher up the pitch than a counter-attacking team. But it should help with finding defenders and midfielders who can break a press by retaining their ability to play out from the back.

So now we can fairly confidently tell, on aggregate level data the type of passes a player typically makes, whether they do this at better or worse than average levels and how well they play under pressure. Plus it all passes the smell test, the names at the top of the list are the names I would expect to be at the top of the list.

But are those names at the top of the list because they are at clubs who are dominant or because they are the best passers? Likely both.

Can we, therefore, find a way of picking out good passers on bad teams and bad passers on good teams using the data that we have?

Other things to consider:

Passing is as much about the options available to you as the mechanics of kicking a ball where you want it to go. If you have great movement ahead of you the options increase as does your chance of making a successful pass, this doesn’t mean you are a better passer.

You also need to consider the skill of the player you are passing to. Some of those Xavi passes hit very fast into feet in a crowded penalty box only worked because they were pinged into Messi, Villa or Eto’o. Most players would not have the deftness of touch to control them. So would Xavi have to alter his passing game if he were on a League 2 side? He would still be the pass master but I suspect it would involve lots more space and he wouldn’t have to demonstrate the same skills.

A pass to feet isn’t always what you want. Busquets often passes the ball into the area he sees space in front of his teammate allowing them to turn and move forward at the same time.

At OptaPro the Barcelona developer gave a  talk about measuring the percentage of time players were orientated to receive a pass. This would be a good option for seeing which players were making themselves available to receive a pass and whether players were spotting this when selecting a pass. It involved full panorama cameras so isn’t suitable for match data but could be useful for training games as a coaching point.

Failures

With passing data, I think we look too much at the successes. Could we learn more about a player by looking at the failures?

Defenders and deep midfielders tend to be passing the ball 50-100 times a game generally with an 80-95% success rate depending on the style of football they play.

Most of these passes seem to me to be neutral events. They are more about retaining possession in deep areas. The thickest lines on the @11tegen11 pass maps tend to be at the base of the team, between defenders and deep midfielders.

These successful passes under little pressure don’t really tell us much.

When we move up the pitch the number of passes reduces significantly. A typical attacking midfielder may make 40 passes and a striker or winger around 20-30 again with a reduced completion rate. Around 60-80%.

This all makes sense, the passes become increasingly pressured the higher up the pitch with less space to operate in.

I wrote before about my “unlucky son” metric. This is based upon the idea that not all failures are equal. There are positive failures and negative failures.  A ball flashed across the 6-yard box that misses the toe of a striker by millimeters is an incomplete pass, just as is a ball passed directly to an opponent who runs through unopposed to score against you.

But we’d much rather the former than the latter.

So my proposal would be that instead of looking at the successful events data which is already captured with through balls, key passes and successful pass ratios we relook at the smaller pool of failed passes probably through video.

Why did they fail?

As far as I can see you’ve got:

Good idea, well executed, unlucky – this would be the defence-splitting passes that just miss.
Good idea, badly executed – this would be the spotted the run but over/under hit pass
Bad idea, badly executed  – pass with little chance of success and no clear plan

We won’t see” Bad idea, badly executed, lucky” as they’ll show up under successful passes.

You would then have the following data.

Number of passes a game / direction of passes
Success %
Expected passes ratio
Difficulty of passes
Passes under pressure
Positive failure ratio

I’ve picked out Ross Barkley as an example of a player who has gone from a high risk / high potential reward passer to a low risk /  low reward passer increasing his completed ratio by 10% but at the cost of all his creative skills. His failures now tend to be fewer but so does his successful/unlucky ratio.

Could this collection be automated?
Potentially.

With tracking data you could measure the options available to a player in control of the ball, the pressure they are under and possibly the best available option to them based on a study of space and movement. Stick in some machine learning and some Markovian stochastic (out of my depth) processes and you’ll get there eventually.

I don’t think we are quite there yet but it is getting closer.

But for now, operating under a hybrid data/eyes method I think looking at the failures is as important as the successes when assessing passing ability.

.