Questions from the analyst desk

It is the use of analytical systems—including statistical methods and, where useful, machine learning—to organize match evidence and describe how an opponent behaves in particular conditions. The goal is to support questions and preparation, not to replace replay review or coaching judgment.

A model can estimate likelihoods based on the data and conditions it has learned from. It cannot guarantee an action, know every hidden cue, or assume that a team will repeat an old choice. Treat a forecast as conditional and update it when new evidence arrives.

Start with permitted replay material, a narrow question, consistent event definitions, and enough context to compare situations. Record the relevant state and timestamps, preserve counterexamples, and make it possible to trace a conclusion back to its clips. More data is not automatically better if its labels or selection are inconsistent.

There is no single threshold that works for every question. Consider the number and diversity of examples, whether they are independent, the number of counterexamples, and whether the sample resembles the upcoming situation. A consequential response should require stronger evidence than a low-cost reminder to watch for a cue.

Use it as a prompt for a flexible branch: a cue to observe, a proportionate response, and a sensible fallback. Explain the confidence and limits to players. Avoid presenting a model score as a command, because the team still needs to read the live match.

Separate discovery data from evaluation data, preferably using later matches for a future-facing test. Compare against a simple baseline, inspect calibration and error types, and check performance across relevant contexts. Also ask whether the output improves the actual review or preparation process.

Usually not by itself. A model may reveal an association between a situation and an action, but it cannot establish motivation or causation from behavioral data alone. Keep observed events distinct from proposed explanations and seek independent evidence for claims about why.

It may reduce how relevant earlier data is, but does not make every historical observation worthless. Record the ruleset, roster, and time period; then review whether the behavior persists under the new conditions. Old evidence should be weighted according to how closely it matches the current setting.

Record what happened and compare the situation with the report’s stated conditions. A counterexample may be an ordinary exception, a sign of adaptation, or evidence that the original claim was too broad. Do not force the action into the old explanation; update the read after the match.

No. A small team can use a modest, carefully maintained notebook and a few well-chosen replay comparisons. The scale of the method should match the available time, data, and decision. A clear manual process is often more useful than an opaque system no one can audit.