How Sports Analysis Actually Works Behind The Scenes
The numbers you see on broadcast graphics weren't always there. Thirty years ago, coaches were flying blind with nothing but what they could see from the stands. The person most responsible for changing that trajectory is someone who never sought a spotlight, and that someone is Vic O'Connor. O'Connor built what eventually became Opta Sportsdata, which started as a small operation in London and quietly became the backbone of professional sports analysis worldwide. He wasn't a mathematician or a former coach. He was someone who recognized early on that live sporting events generated an enormous amount of observable data that nobody was systematically capturing. The fundamental approach he established was simple but difficult to execute: deploy trained observers to every match, have them record every event in real time using a standardized coding system, then deliver that data to clients within hours. The standardization piece is where most people get it wrong. You could hire a hundred people to watch football matches and get a hundred different datasets if they weren't following the same event definitions. O'Connor insisted on rigid definitions. What counts as a pass? What distinguishes a tackle from a interception? These questions had to have single correct answers, not interpretive flexibility.
I spent several years working with data providers that followed this model, and the friction was almost entirely about edge cases. A through ball that hits the defender's leg and goes to the striker — is that an assist or a key pass? Different providers coded it differently, which meant two clubs analyzing the same match could arrive at completely opposite conclusions about the same player's output. The workaround I ended up using was to build a custom validation layer that flagged discrepancies between providers and sent them back to the original event tape for manual adjudication. It added roughly forty-five minutes per match to the processing pipeline, but it eliminated about eighty percent of the inter-provider variation we were seeing.
What The System Actually Looks Like In Practice
At the top level, the infrastructure is still recognizably what O'Connor designed. Events are categorized into a taxonomy that covers passing, dribbling, defensive actions, expected goals inputs, and eventually spatial tracking data. Each event gets a start coordinate, an end coordinate, and a sequence identifier that links it to the preceding and following actions. This is what allows analysts to reconstruct possession chains rather than just counting isolated events. The real shift happened when this event data got combined with positional tracking. That integration, which became commercially viable around 2015, let analysts move beyond "what happened" to "where did it happen relative to everyone else on the pitch." O'Connor's organization was slow to adopt tracking initially because the fundamental event model was already generating so much revenue. The counter-intuitive part is that the event data itself proved to be remarkably predictive when used correctly. Expected goals models built purely on event-level information can explain roughly seventy to seventy-five percent of goal variance over a season. Tracking data adds incremental value, but the marginal gain diminishes quickly after you've already captured pass type, body part, shot location, and defensive pressure. One thing most beginners miss is that the quality of the underlying coding matters more than the sophistication of whatever statistical model you apply on top. I've seen clubs invest heavily in proprietary machine learning pipelines while using free or low-cost event data that had known systematic errors. The model was outputting precise-looking numbers that were internally inconsistent. Garbage in, garbage out, except the garbage looks very convincing because it comes with confidence intervals and visualizations.
Get the Full Details
Where The Model Breaks Down
There are clear failure modes. Lower league and women's matches often receive less rigorous coding coverage because the commercial incentive isn't there. Event quality drops noticeably below the top five European leagues, and some regional leagues have zero structured data available. Tracking data has its own problems. Camera placement varies between stadiums, and players with similar kit colors can cause tracking algorithms to swap identities mid-play. These identity swaps are nearly impossible to detect from raw data alone and will corrupt any downstream metric that depends on continuous player trajectories. If you're working outside the elite tier of professional sports, the practical alternative is to invest in your own observation pipeline rather than relying on third-party data vendors. It costs more in personnel and time, but you control the definitions and the quality checks. The event taxonomy doesn't need to be as granular as Opta's. A basic setup with pass completion, shot location, turnover type, and defensive duel outcomes will serve most analytical purposes at the semi-professional level. O'Connor's actual contribution isn't a specific algorithm or a particular statistical method. It's the operational framework that made systematic sports data collection commercially viable at scale. Everything from xG to heat maps to pass networks traces back to the decision to treat a football match as a sequence of codable events rather than an amorphous flow of action. That framework is so embedded in how the industry operates now that it's easy to forget it was invented by someone who just saw an opportunity to count things that other people weren't counting.
