Redundancy and predictability in natural language in sports analytics

The Weak Side, Two Months Late

The report went out on a Thursday. It named the opponent's left centre-back as a build-out weakness, cited a heat map showing he favoured his right foot under pressure, and recommended pressing him early in possession sequences. The press worked precisely once, in the eleventh minute, and then stopped working. By half-time the centre-back was receiving on his left with no hesitation, playing diagonals the report had not modelled, and the opposing manager had clearly adjusted the build-out weeks earlier — the analytics team simply hadn't caught the change.

What had happened was not a bad model. The heat map was accurate. The tendency was real. It had simply expired. The centre-back had been retrained by his own coaching staff seven weeks prior, after two matches where the right-footed bias had been exploited by a different opponent. The report's author, a performance analyst working from a database of the last full season plus early-season tracking data, had built a game plan on a pattern that was true when it was recorded and false when it was used.

Where the Failure Actually Sits

Call this a stale prior and the diagnosis half-answers itself, but it is worth being precise about what kind of staleness. The analyst's dataset was not sparse. Thousands of frames of tracking data supported the left-foot-avoidance claim; it was not noise, it was a genuinely strong signal — for the period it covered. The failure was not statistical, it was temporal: the model had no mechanism for asking "is this still true," only for asking "was this true across the sample." Nothing in the pipeline flagged the tendency as time-stamped and decaying rather than settled and permanent.

This is the same failure that shows up whenever prediction substitutes for observation of a moving target. Injury reports go stale within days. Transfer activity changes a squad's shape within a window. An opponent's pressing trigger, tuned across a mid-season tactical shift, can invert in three matches. A performance analyst working from historical tendencies is doing something structurally similar to language prediction: inferring the next likely event from redundancy in the past record. The trouble is that redundancy is only ever a measure of what has recurred, and the world sports teams operate in is one of the more actively adversarial environments for recurrence — because the opponent is also watching the tape, also trying to become unpredictable, also correcting exactly the pattern that got exploited last time.

The Concept Underneath

Natural language is heavily redundant: Claude Shannon's 1951 guessing experiments put the entropy of printed English at roughly 0.6 to 1.3 bits per character, against an alphabet capable of carrying 4.7. That surplus — something like 75 per cent redundancy — is not waste. It is what lets a reader restore a smudged word or a listener recover a syllable lost to static. Redundancy exists because the things language describes recur: grammar repeats, collocations repeat, and the world it refers to repeats. Prediction is only possible where something is stable enough to be predicted from.

A tendency report is a compressed statement of exactly this kind of redundancy, applied to a football pitch instead of a page. "Prefers left foot under pressure" is a linguistic compression of thousands of repeated observations, useful precisely because footballing habit is usually stable across a season. Most of the time it is. That is why scouting works at all — most tendencies persist long enough to be worth reporting. The residual, the part no historical sample can supply, is whatever has changed since the sample was taken. And in a competitive league, what has changed since last month is exactly the part that matters for next weekend, because both sides are actively working to make the stale part stale.

Three Positions on the Same Pitch

A Large Language Model, applied to this problem, is a redundancy engine trained on the corpus of past matches: it will recover the centre-back's habitual bias faithfully, because that bias really was stable across the training window. Its blind spot is not a modelling defect. It is the definition of a frozen corpus — nothing in last season's footage can report a change made in week six of preseason training.

A Large World Model closes part of the gap by sensing the current scene rather than reading a report about an old one: live tracking data showing the centre-back receiving comfortably on his left, this match, right now. That is real correction, and it is the reason live tracking overlays have become standard on the touchline. But the correction only lasts as long as the scene is in view. The moment the analyst looks away — between matches, during the international break, across the transfer window — the moving part is unmeasured again, and the next report risks freezing a habit that has already moved on.

A Large Universe Model is the position that keeps every relevant stream running continuously rather than sampling it and freezing the sample: tracking data, injury bulletins, transfer activity, opponent tendencies, all held as beliefs with a timestamp and a source, revised the moment a stream contradicts them, decayed in confidence the longer they go unconfirmed. Not a smarter model of last season. A standing watch over this season, updated as it moves.

Two Objections Worth Taking Seriously

Attach a live feed to the existing database and the problem is solved. That's an integration, not a new category of system.

Partly right. A club that pipes live tracking data and injury updates into its analytics dashboard has genuinely improved on the static report, and this is exactly what better-run analytics departments already do. The limit is that a feed answers "what is happening" only for whatever it happens to be pointed at. It does not, by itself, tell an analyst that the left-foot tendency was last confirmed 51 days ago and has a decaying confidence score, versus a pressing trigger confirmed in the warm-up eight minutes prior. Piping in more data streams is necessary. What is missing without deliberate design is provenance and decay: knowing not just the current reading but how old each belief is and how much to trust it. That is a design commitment, not an automatic consequence of adding a feed.

More streams means more noise. Tracking vendors disagree on player load, injury reports from different sources contradict each other, transfer rumour and confirmed activity blur together. Continuous intake multiplies the reconciliation problem it was meant to solve.

This is the real cost, and it should not be minimised. Adding sources converts a prediction problem into an estimation problem — which source, how recent, how reliable — and that work is genuinely harder than running a season-end regression. But the same redundancy that helps in language helps here: three independent tracking feeds that disagree bound the truth better than one confident but outdated scouting report. The reconciliation problem is real work. It is also more honest work than pretending last season's heat map is still current, because it produces a system that knows what it doesn't know, rather than one that is silently wrong.

Why This Is the Top Rung

None of this makes tendency reports worthless, and none of it promises a system that never fails. What it changes is the shape of the failure. A frozen corpus fails by being confidently out of date. A live scene fails by going blind the moment attention moves elsewhere. A system built on every stream, still running, with provenance and decay attached to each belief, fails by flagging its own uncertainty rather than concealing it — the press recommendation would have arrived with a note that the underlying tendency was seven weeks stale, not a false certainty. Past that point there is no further category of intake to add. There is only more of it, checked more often, weighted better. That is the terminal rung on this particular ladder — not the end of what analysis can do, but the end of what "more intake" can buy.