Line go up, hopefully

'Line may go down as well as up' is a pretty common bit of financial advice. Reductive as it may be (some financial types in the UK reckon that it's why Brits don't retail-invest much), it's a lot snappier than 'line may go up slowly or quickly, or down slowly or quickly, and may go up or down in ways different to how it's been going up or down lately'.

Rate of change isn't linear. You'd think that humans would have a better intuitive understanding of this: every day begins with darkness before a rapid increase in brightness. Will the brightness continue increasing exponentially?! Yet no, the brightness plateaus.

To be fair to our species, it's incredibly difficult to tell when the rate of change will shift. (Unless you're talking brightness in the north of Britain, which often just peaks midway through dawn). You see this all the time with young footballers: they burst onto the scene, their rate of improvement is remarkable, but where their performances settle varies hugely.

You could model this, but this isn't actually a post about modelling that. It's a post about robo-data.

Four years ago, a different set of kids were on the rapid growth curve: "The kids [of football analytics] are growing up, putting on business suits, and going out into, and shaping, the big wide world." That was a line in this newsletter, which also might've been the earliest outing of the Get Goalside schtick about football clubs doing tech: "(A new Warholism for you: "In the future, everyone will be a tech company for 15 minutes")". But, more significantly, it raised the idea of doing away with your event data subscription and just generating the data from tracking data instead.

Last week, this post floated into my LinkedIn feed: "You can now use AI to analyse an entire 90-minute football (soccer) match for under $1 on SPAN, using Gemini 3.7 Flash". Now, there are layers of Get Goalside caution here: Firstly, 'analyse an entire match' corresponds to 'Get AI-tagged corners, free kicks, interceptions, goals, shots on target, shots off target, entries into the 18-yard box and throw ins'. Dampen the expectations. Secondly, the company's previous post on benchmarks, using Gemini 3.6 Flash, wasn't hugely impressive in its results (although refreshingly honest).

And then there's this recent paper which suggests Vision-Language models are often picking up non-football indications of football actions (e.g. TV overlays), rather than picking up the relevant on-pitch action. And not only can you use other computer vision and machine learning methods, rather than multi-modal models, but FIFA is on the case. FIFA has four things this kind of work needs: access to event data, access to video, resources, and a track record of following through with things they put their mind to.

Where in the curve of progress are we with automated data generation? What is the current level; what is the rate of change?

Elsewhere in that 2022 post, I wrote "[I]t would seem weird if the increasing accessibility of computer vision programmes didn't end up affecting relationships between data providers and clubs (or other data purchasers) in some way.". That change seems to have happened more in the amount of companies that now collect and offer broadcast-footage tracking data, rather than clubs doing it themselves. (In part because clubs are spending their time bringing boring data storage and applications in-house, instead of doing fun stuff, although I do understand why). Or maybe the ones doing this are just quiet.

That part in parentheses is a classic example of why predicting the rate of change can be hard. Any prediction about future trajectories rests not just on one assumption, but many, and each of them could affect the future path in their own ways. A youth player might be on the rise, and may be able to see off challenges of poor teammate quality, manager upheaval, a major injury, but might be knocked off-course if that injury becomes persistent.

Attempts at auto-data generation have various issues. One of them is that, in football, the ball is very very small on-screen. So if events can be detected only through player trajectories, that affects the range of possible paths open to this particular rate of change. Big if. But an existing if.

Line may go down as well as up. Availability of video can be clamped down on, investment bubble can pop, professional footballers may quit the sport en masse for padel.

I hope line goes up.