There's a distinction in football data between events and tracking. Events are things that happened — a pass here, a tackle there. Tracking is the position of every player and the ball, continuously, several times a second.
Tracking is vastly more informative and until recently it was vastly more expensive. That gap has closed, and the closing is probably the most consequential technical development in football analytics in the last decade.
The old economics
Traditional optical tracking required a permanent camera installation at the ground — multiple synchronised cameras, calibrated, with the processing infrastructure behind them. Some systems used wearable devices instead, with their own installation requirements.
That meant tracking data existed only where somebody had paid to install a system, which meant top-tier leagues and not much else. If you wanted to analyse a player in the Portuguese second division or a Brazilian state championship, you had event data at best.
The practical consequence was a two-tier analytics world. Rich leagues had rich data. Everywhere else was assessed with cruder tools, which meant the analytical edge in recruitment was concentrated exactly where the money already was.
What changed technically
The problem is extracting 3D pitch positions from a 2D broadcast feed shot by a moving, zooming camera that follows the ball and frequently doesn't show most of the players.
Several hard sub-problems had to be solved. Detecting players reliably against varied backgrounds and crowd noise. Tracking them across frames including through occlusions when players overlap. Re-identifying a player after he leaves frame and returns. And camera calibration — working out, frame by frame, exactly where the camera is pointing and at what zoom, so image coordinates can be mapped to pitch coordinates.
Each of these has been substantially addressed by deep learning approaches over the past several years. Pitch calibration in particular benefited from the fact that football pitches have known dimensions and visible line markings, which gives a strong geometric prior.
The remaining fundamental limitation is that broadcast cameras don't show everybody. At any moment several players are out of frame, and their positions have to be inferred. Approaches vary — physical plausibility constraints, learned models of typical team shape — and the inferred positions are considerably less reliable than the observed ones.
What it enabled
The applications that tracking unlocks are qualitatively different from what event data supports.
Pitch control and space models. Given everyone's position and velocity, you can compute which team would reach any point on the pitch first. That gives you a continuous map of controlled space, which is a far better description of dominance than possession percentage.
Off-ball evaluation. The runs that don't get the ball, the positioning that closes a passing lane, the defensive movement that prevents an attack from starting. All of this is invisible to event data and central to football.
Option-aware passing analysis. Instead of "he played sideways," you get "he played sideways when a forward pass to a teammate in a higher-value position was available with a given completion probability." That's the difference between recording a decision and evaluating one.
Physical output in context. Distance covered is more meaningful when you know where and in relation to what.
The recruitment consequence
Cheap tracking in previously unmeasured leagues is the change with the largest financial implications.
If you can build a proper positional model of a player in a league that nobody else is analysing rigorously, you have exactly the informational edge that analytics was supposed to provide and had largely lost through universal tool adoption.
Several clubs have reoriented recruitment around this — targeting markets that are under-analysed rather than under-scouted, which is a different thing. The scouting was already happening in those leagues. The systematic data wasn't.
The caveats worth holding onto
Some honest limitations.
Broadcast-derived tracking is noisier than installed-system tracking, particularly for off-camera players, and any analysis has to account for that. Comparing results between the two sources is risky.
The pitch calibration depends on visible line markings, so it degrades in poor weather, on worn pitches, and in unusual lighting. Lower-league footage is exactly where these conditions are most common, which partially undercuts the main use case.
And the models built on top inherit all the usual problems. A pitch control model has assumptions about acceleration profiles baked in. A passing-option model has assumptions about what players can see. These are reasonable assumptions and they're still assumptions.
Where it's heading
The obvious direction is towards full automatic reconstruction — not just positions but body orientation, head direction, and eventually gaze. Body orientation is already being extracted with reasonable accuracy and it's enormously informative, because a player's shoulder position tells you what he can see and what he's prepared to do.
Head direction would let you actually measure scanning, which has been discussed in coaching for years on the basis of a small number of manually-coded studies. Being able to measure it at scale would settle a lot of arguments.
That's coming. Whether the sport does anything sensible with it is a separate question, and the record on that is mixed.