Ian Graham was Liverpool FC’s director of research from 2012 to 2023, charged with applying data analysis to the club’s transfer policy. During this period, the club was a trendsetter in the use of data. Graham, who holds a PhD in physics from Cambridge, is now chief executive of Ludonautics, the sports advisory business that he founded. Its clients include football clubs.
Simon Kuper: Clubs such as Brighton, Brentford and Liverpool under John Henry’s Fenway Sports Group have benefited from using data analysis. But few others have followed their lead. You have written: “Even today, most Premier League clubs do not take data analysis seriously.” Why is that?
Ian Graham: It’s an organisational problem, and it’s a people problem. Every owner, and I put FSG in the same basket, says “We’re going to come in, we’re going to do things professionally.” But me saying that, and you actually carrying out my orders, are two very different things in a conservative sport. The easiest way to get a new job in football is to be in the middle of the pack. Now, at the top end that really doesn’t work, but most people aren’t at the top end. If you do something different and fail, you’re marked with that for the rest of your career.
It is changing. The younger generation of managers buys into this in a way that the older generation doesn’t.
SK: So use of data is spreading?
IG: It’s slower than it may appear from the outside because people have a motivation to tell you how smart they are. The data people have a lot of information, they even have sophisticated people analysing it, but it doesn’t really inform decisions.
There’s so much low-hanging fruit that just rots on the vine. Most clubs still don’t pay due diligence on transfers. Not even complicated data analysis, but simple data analysis.
SK: You’re presenting two different narratives. One is that the game is getting smarter. The other is resistance to change.
IG: Yes, there are two narratives. For all I complain about the slow adoption in the Premier League, it’s pretty good compared with the rest of the world. Spain in particular is pretty slow and traditional. Within the Premier League, that change is currently halfway through happening. Manchester United have not made that change yet. Chelsea didn’t have to work that way, so now they’re going through the difficulties of that change. When I started at Liverpool it really was just Liverpool doing it. Arsenal were doing it behind the scenes but it wasn’t really having an impact.
SK: Most clubs still seem to place more importance on choosing their head coach than using data.
IG: It’s still the standard approach in Italy, in Spain, and those teams still do very well in [mainland] Europe: the traditional coach-led show, it’s all about the tactics, that’s what’s going to win the game.
The number of coaches who make a difference in that way is limited. That approach puts all your eggs in one basket, the coach basket. It’s similar to the number of transfers that fail. You can do exactly the same thing with coaches. You can say, “We’re going to hire the next Guardiola,” but what are your chances of success?
If coach-led means the coach dictates recruitment, it’s limited. The coach has got two games a week to coach, he has no time to do recruitment. We spoke about [Erik] ten Hag [at Manchester United] recruiting Dutch players because that’s who he knew, and players in the Dutch league. They recruited good players, just not for a good Premier League club. Understand what you’re getting from the coach. You’re not getting 10 points for free from most coaches.
SK: Football clubs traditionally functioned as autocracies. Either the coach decided everything, or the club owner tried to, usually based more on whims than on data. But Liverpool and Brentford have committees to decide transfers. Surely football will move that way, towards the wisdom of crowds?
IG: The owners are always great for sticking their oar in and making a renegade decision. “What’s the point in owning a football club if I can’t sign my favourite player and I have to listen to what some spreadsheet says about data?” I think the coach as dictator is specific to England, and is something that is dying off. It is more wisdom of crowds.
SK: Still, it seems that Liverpool’s manager until last season, Jürgen Klopp, achieved such status towards the end of his tenure that he acquired great personal power.
IG: Jürgen created a lot of success for the club, so it’s understandable why it moved in that direction. I’m happy to talk about my colleagues persuading Jürgen [in 2017] that Mo Salah was the player to buy instead of Julian Brandt. In 2022, he signed Darwin Núñez [for £70mn plus add-ons] instead of Alexander Isak. Both players, if you look at top young centre-forwards in Europe, they would be number one and two — or two and three but [Erling] Haaland was going to [Manchester] City and out of our price range. Jürgen preferred Núñez. It would be very churlish of me to say, “It’s terrible that Jürgen had his choice”, when in the past Jürgen had been persuaded by me and my colleagues of a different choice. And it was still the case that we signed good players — in Núñez’s case, one of the best young strikers in Europe. [Nuñez, now out of favour at Liverpool, is reportedly about to be sold to Napoli for around half the sum Liverpool paid for him.]
SK: For about 20 years now, clubs have had event data, which measures what players do on the ball: chiefly passes, shots, tackles. Now we also have tracking data, which measures their movements around the pitch. What new insights does that provide?
IG: From about 2016, tracking data where you see 25 frames per second, 22 players, became available. The amount of resource that’s required to extract insight from that data is an order of magnitude more. That’s where the cutting edge is. The next level of data is that instead of one point per player, 25 frames per second, we get to 29 points per player, 25 frames per second, which gives the location of all their joints, their eyes, various way-points on their head, and so on. So you can see: this player’s doing a jump, this player’s making a kicking action, this player’s looking behind them. You don’t currently get that from tracking data. Very few if any clubs will have the resources to even ingest the data.
SK: Will tracking data help clubs better evaluate defensive players, who do most of their work off the ball? That might raise their transfer fees towards the level of attacking players.
IG: The most difficult position to analyse is centre-back. They’re the most off-ball. So Ibrahima Konaté [signed by Liverpool from RB Leipzig in 2021] had a much higher rating in our tracking model than in our event model. Why? Because Leipzig plays suicide football and he has to defend half a pitch five times a game. So the event model is saying, “So much danger is coming through Konaté.” The tracking model can see, “Well, he’s literally the only player in his zone”.
It leaves us a more sophisticated understanding of what matters and who’s good. You can ask, “Is this player consistently picking what we consider to be the high-value option?” The event data can’t see the other options. Tracking data lifts the lid on it all and says, “This is exactly what the player had in front of him.”
SK: How might AI change data analysis?
IG: The public conception is, “I’ll feed all this data into my AI model, and my model will identify the next Lionel Messi.” That’s wrong. The way it makes a difference is certainly with data collection, so tracking data is computer-vision algorithms, which are AI algorithms really. We did a paper with DeepMind [Alphabet’s AI research laboratory] looking at predicting where players would run in the next 10 seconds. We’re pretty good for 10 or 15 seconds, until the next pass or the next challenge between two players happened.
I dislike AI because it’s a black box. It’s getting better, but it finds it difficult to explain its reasoning to you. When you look at tracking-data models, you start with a simple model, where if you ask me, “Why does your model say this?” I can pull apart each piece and say, “This is why.” But it’s all simplified, because football is such a complicated game, so our ambition is to start taking out those simple pieces, put in an AI model that is less explainable, but you can see the differences it’s making to the model. You can see what your model is missing, and what the AI is better at predicting.
SK: What questions might a club analyst ask AI in two years’ time?
IG: The big difficulty is to say, “In our [club’s] system, how would this player perform?” You can use simple models to look at how many times you get on the ball in your current team versus what you’ll get at our team, how many players will you have in front of you. But I can imagine AI will do a better job than a handmade model.
SK: So in theory, AI could take somebody who’s playing for Lille and ask, “How would he look at Manchester City?”
IG: We currently try to do that, but AI can help give a more tailored answer to that question.
SK: What changes do you expect in the use of data to recruit players and shape tactics?
IG: Recruitment’s the big-ticket thing that clubs do, so it’s the most important thing to get right. When I was at Liverpool, the most impactful thing you could do was sign good players and pray that they’re going to make the impact on the pitch, because the coaches are the experts when it comes to tactics. It’s hard to make a difference there.
But today, the new generation is [asking], “You can say something about recruitment, why can’t you tell me something about my game? Tell me where my position is going wrong.” It was interesting talking to this centre-back. He said, “Even if you’re wrong, I’m interested to see how you think of my game. I’m interested in my next transfer. What makes me look good and what makes me look bad to data people?”
SK: Which single type of data on players that currently isn’t collected would you like to have in future?
IG: The mental side of the game is super important and very difficult to measure. Well, you’ve got a whole host of data sources directly from the players through their social media and interviews, and so on. There may be some clues, or you can look at differences between their normal statements if they’re always positive — maybe they’re slightly less positive. Taking large quantities of unstructured, difficult-to-process data such as body language and facial expressions in interviews, that’s the sort of thing that AI is good at.
We pitched this as our first idea to DeepMind and they decided against doing it, I think on either moral or ethical grounds. It comes down to that ethical issue: should you be doing this? I understand why they got cold feet.
This transcript has been edited for brevity and clarity