Simulation
A closed model versus a match that reacts to being predicted. More data narrows the first and cannot fix the second.
A statistical model can tell you that one side has a 57% chance of progressing. That is a genuinely useful statement, and it is not a prediction that they will win — it means that under the model’s current assumptions, the other side progresses in roughly 43 out of 100 comparable situations. By the time a headline has converted it into “AI predicts England will beat Mexico”, the probability has become a promise, the assumptions have disappeared, and the limits of the data have been quietly discarded.
This simulation puts the two things side by side: a closed model that behaves exactly as specified, and a dynamic match that does not.
The distinction matters more than the arithmetic. A closed system can, in principle, be characterised completely: gather enough data, and residual uncertainty shrinks toward zero. An open system contains intelligent participants who observe conditions and adapt — including adapting to the forecast itself. A side described as overwhelming favourites may become cautious and protective. An underdog may take risks precisely because conventional play offers so little chance. Managers read the same analysis and make late decisions specifically intended to invalidate the opposition’s expectations.
That is reflexivity, and no quantity of additional training data removes it, because the act of predicting changes the thing being predicted.
Run the closed model first and let it settle. It converges, because it is a system whose rules do not react to being modelled — the only limit on accuracy is how much you know.
Then run the dynamic match. Watch how the outcome distribution refuses to collapse in the same way, even though the underlying quality difference between the sides has not changed. Deflections, refereeing decisions, injuries, fatigue, weather, crowd behaviour, and choices made under pressure all enter as the match proceeds, and each one alters what follows.
The useful habit is to stop asking which side the model favours and start reading what the spread is doing. A model that reports a narrow, confident range for an open system is not more accurate than one reporting a wide range — it is more poorly calibrated.
Uncertainty is routinely discussed as a defect that better technology will eventually eliminate. In football it is closer to being the product: nobody stays up until 3am to watch the statistically superior plan execute itself inevitably.
For product work the same trap appears whenever a forecast is presented as a single date or a single number. A useful prediction states the most likely outcome, the confidence in it, which variables dominate, what assumptions are in force, and what would invalidate it. That is less dramatic than a clean answer, and considerably more honest — the same argument the probabilistic forecasting article makes about delivery dates.
Press play to watch reality aggressively introduce itself to the statistical forecasts.
Change one input at a time, run the model, and compare the result with the starting state. Then repeat the experiment with a different constraint or strategy so you can see which relationships drive the outcome.
Look for trade-offs, thresholds, feedback loops, and points where a locally attractive decision produces a worse system-wide result. The simulation is intended to make the article's idea observable, not to predict a real operation.
This is a deliberately simplified model. It omits the data quality, exceptions, human judgement, and operational constraints of a live system, so treat its behaviour as an illustration of a mechanism rather than as a planning recommendation.
Read: Chris Kamara, AI, and Why Football Refuses to Be Predicted →