Mislabeled Data: The Silent Flaw Inside Professional Football Analytics Pipelines

In August 2026, at SportsData Korea, I sat in front of a dataset covering 142 K League 1 matches played without spectators. The task looked simple: compare it against 142 pre-pandemic matches to measure what losing the crowd actually did. Filtering the "pressing sequence" field, I found 41 sequences labeled as pressing that in fact began from a throw-in or a corner. The model downstream learned from them, and learned wrong. Nobody caught it for three months, because a bad label always looks reasonable: a pressing sequence off a throw-in still carries intensity, speed and coordinates. It fails on one point only — the team was never in a state of free loss of possession. Gaps do not disappear on their own; they simply change their name to failure. Professional football analytics runs on a multi-layer data supply chain. Providers such as StatsBomb, Opta and Wyscout collect events; labeling teams across several countries classify them; analysts convert them into metrics; coaching staffs turn metrics into session plans. Every layer depends on the one above it, and the first layer — domain labeling and event labeling — receives the least auditing. In the K League, most clubs keep only two to four in-house analysts. They do not have the manpower to re-check every event delivered by a provider. They take the package, trust the label, build the session. Many V.League clubs follow the same model on thinner budgets. Errors at the labeling stage therefore travel further than the end user can detect. Inside that same 2026 dataset, the home win rate fell from 47% to 41.5%, and average goals per match rose by 0.7. Those findings only hold if the labeling layer above them is clean. What matters here: labeling errors rarely produce absurd output. They produce plausible output pointed in the wrong direction, and the material used to build the model was contaminated before the algorithm ran its first line. There are three labeling layers, each with its own failure mode. The first is the domain label: does the file, article or report actually belong to the field being analyzed. This is the most neglected layer, because it is assumed that anything sitting in a football folder is football data. A general dataset can easily contain medical, traffic or weather records. One wrong tag at this layer and the entire downstream pipeline processes material outside the football domain as if it were tactical data. The most serious consequence sits in a model predicting wrongly with complete confidence. The second is the event label: what happened on the pitch. This is where tactical analysts work daily, and where the most subtle errors appear. A counterattack starting from a throw-in gets merged with a counterattack from free open-play transition. A penalty-kick shot is folded into an open-play xG model. A possession sequence by a team a man down is fed into training material about possession at numerical parity. In late June 2026, reviewing the tape of South Korea against Germany, I counted 87 deliveries into the box by Germany and only two shots on target. Count by the "final third entry" label used by some providers and the result shifts, because corners and live play are merged together. The difference lives in the definition, not in the raw numbers. Between two passages of play, time exposes the decisions the eye skips over. The third is the context label: match state. The same action carries entirely different value depending on scoreline, clock, players on the pitch and tempo. At the 2026 World Cup, across six matches, Morocco shifted into a 5-4-1 with an average of 2.3 seconds measured from the moment of losing the ball. Full-back Achraf Hakimi advanced an average of 58 metres per match, and when he dropped, the flank space was covered by midfielder Azzedine Ounahi. A model that tags every deep Morocco block as defensive will miss that this team defended with structure, and that structure was built before the ball reached the opponent's feet. The industry's usual response to discovering bad data is to buy more data. Clubs extend provider contracts, add events per match, layer in skeletal tracking. That approach assumes the problem is volume. The problem is verification. A domain check gate placed at the entrance to the model costs far less than an expanded data contract, but it demands something professional football rarely has: the patience to say there is not enough information. Rejecting an obviously off-topic file is easy. Rejecting a file that looks relevant while sitting outside the football domain is harder, because it breaks the founding assumption of the whole analysis group. A provider's reputation does not protect you; it only tells you where to look first. The largest blind spot in modern football data analysis is that a bad label never raises an alarm. It triggers no compile error, collapses no model, lights no red warning on the dashboard. It only sends a coaching staff to prepare for something that does not exist on the pitch. Data means something only when we ask at the right moment; ask at the wrong one and every figure is noise. Through the annual season, as each matchday generates hundreds of thousands of event records, the pressure to publish fast will keep pushing post-hoc verification to the bottom of the priority list. Based on my experience tracking matches, the cheapest check remains this: pick the single metric your team uses to prepare for the next fixture, trace it back to the original record, and ask whether its label matches what actually happened on the pitch. If the answer is no, the model was never the problem.