Forgive me, for I have been watching Drive to Survive - the Netflix Formula One series - for the first time. And like a new convert, I'll be talking about it.
The amount of footage and array of angles on offer make for an exciting show alone, never mind the personalities and storylines. You really get the sense of how fast the cars are going, how they stick on the track and cannon into corners almost sideways in a way that cars simply shouldn't.
That they do is all because of the downforce that the meticulously-engineered cars generate, quite often only possible because of the high speeds they're moving at. Pushing too hard can cause trouble, but easing off the accelerator doesn't necessarily make things safer.
And so we get to football analytics.
In 2017, a group of analytics researchers released a piece of work on defensive 'ghosts'. It's like creating a FIFA AI from real data, that reacts to developing action just like real players would. (Click here for the full ghosting paper, or see this link for a 'Research in Focus' handy summary).
They also released a short, two-and-a-half minute video. In it, they highlight two examples where the ghosts, which have been trained in a model to be how 'league average' players might play, act more proactively than the real-life defenders actually did.
In the second example, attention is drawn to a team sitting back as a collective unit. This reminded me of one of the earliest ideas I can remember seeing analytics folk advocating: stop sitting back when you go a goal up.
Stop easing off the accelerator.
It's something that is referred to in Christoph Biermann's book Football Hackers: "[Matthew] Benham's stats had convinced [FC Midtjylland manager Glen] Riddersholm that a narrow lead was best defended by going aggressively for another goal; playing not to concede, on the other hand, only decreased the probability of scoring and and increased the opposition's chance to find the net." (chapter 3, Rise of the Outsiders).
It might intuitively feel more comfortable to be less aggressive in some circumstances, but, just like in F1, the best approach might actually be to keep the foot down.
By chance, a few days after I recently revisited the 'ghosting' video a football match I was watching featured a turnaround by one team, partly brought about by a more aggressive approach in their defending after half-time. Their opponents were suddenly getting no time or space, the newly-aggressive team won the ball back quicker, they were able to assert themselves on the match.
It's hardly limited to that match either, it happens quite often. So why is the actually-slightly-counterproductive retreat such a common feature of football? And what can analytics do about it?
I think there are three things at play, two which have particularly interesting and important implications for making analytics models.
The first and most simple is that sometimes it works. Football is a low-enough scoring sport that it can be possible to just ride out a match while defending deeply or passively. (That said, I think the most common version of this is that a team defends this way for about 20-30 minutes and then realises that they have to move up a gear again).
That shouldn't matter too much in modelling though: if it works then the model would recommend it. It's the second and third that are trickier; two F's: fear and fatigue.
The reason why players defend passively is not that they're saying "ah, winning the ball back isn't important", it's that they're saying "we're worried they'll get to the space in front of goal, so we'll give them space further away from goal instead". It's a tactic of fear.
Sometimes fear is sensible. But sometimes it's just a lack of confidence. And in those circumstances, though easier said than done, the best approach is to get confident.
The thing that modelling might struggle to help you with is knowing when the fear is justified. When a defender goes to tackle a forward as they receive the ball on the halfway line, the margins are tight. If the challenge doesn't go to plan, they might need to turn and sprint back quickly. Nobody enjoys that.
Part of the reason no-one enjoys that is fatigue, the second f from earlier. Players get tired, and that affects their speed covering distances and accelerating. Any model that relies on these movements as factors (such as ghosting or pitch control) might also need to consider fatigue. A model that is 'naive' to tiredness might suggest that a player is in the wrong place; the player themselves might say that they were stood where they were stood because they couldn't recover as normal.
Fortunately, although it might be hard to factor this into analytical models, it shouldn't be impossible (even if we don't hear about it in public for a while). [There might already be public research relating to this but I don't have it easily to hand]
For pitch control models - which work on the basis of how long it would take for a player to reach a particular point of the pitch from where they are now - player speed can be built into them. It stands to reason that a fatigue factor could be too. Maybe this would be based on time spent on the pitch, or distance covered at a high-intensity.
Although each individual will tire differently, it seems likely that there'd still be some kind of general rule that you could follow. A halfway house between a complete average and an individual approach could be positional group: research shows what we all expect, that different roles have different physical outputs (a recent paper that touches on some of this position-to-physical metrics interaction is here).
If you really trusted the quality of your tracking data, you could even help to manage defenders' fear levels in-game. The next few sentences are a bit of dreamworld, but why not live in dreamworld for a little while?
Let's say you build a profile of an opposition forward with their levels of acceleration and speeds they can hit in matches. Then you monitor their reactions in the early part of a game and see where in the scale their acceleration and speed levels are falling. If they're below their regular levels, then you could tell your centre-backs to have more faith in getting touch-tight.
Going even further, and putting even more faith in the tracking data and modelling ability, maybe you could measure reaction times. Maybe you could run a defender's movements through a model and it could say "they've reacted slower than an average defender would to these interceptable passes and these off-shoulder runs". And maybe you could take that and tell the player that maybe they should either focus a bit more or to ease away from the tight challenges.
I don't think that tracking data is at the level where it could make that possible, but advances are being made all the time.
It seems almost unnatural to be able to coach fear out of people in circumstances where it would be quite reasonable. But sometimes in sport it really helps in winning.
In one episode of Drive to Survive (season 3, episode 8), driver Carlos Sainz is shown graphs, which I presume show different levels of forces the car was experiencing while going around corners. His lines were being compared to his teammate, Lando Norris's.
"So, compared to Lando, this was his fastest through turn one compared to our fastest through turn one," someone who I assume is some kind of engineer says.
"You can see he's challenging the fronts [front tyres?] just as much in this phase," a biro points to the laptop on-screen, "but his style is that he won't extend holding the brake here because he always just programmes himself to come off brakes the same each corner. It's just the subtle difference that he's got the tow[?] here, you've not quite got it."
Imagine being able to break down a video clip of a defender challenging for an interception-duel in a similar fashion. "You've reacted to the opportunity quickly, but the speed of acceleration is slower [than your teammate/than the opponent/than your average]. That was the difference between making the tackle and committing a foul."
We can dream.