Prospective tracking: predictions frozen before kickoff

Almost any model looks impressive when it is tested on games whose results are already known. The honest test is the one that happens in order: the prediction first, the game second, the grade third. That is the only way to know what a model would really have told you at the time.

1. The pregame read is created. The model runs on information that existed before the game — no result, no box score, nothing from the future.
2. It is frozen. The output is recorded with the time it was made. From that moment its value, its inputs and its timestamp do not change.
3. Kickoff. Only a record made before kickoff counts as a pregame prediction. One made at or after kickoff is kept apart and never counted as one.
4. The evidence arrives. After the game: the final result, the official statistics, and the record of which players actually took part.
5. Settlement joins the two. The result is attached to the ORIGINAL record — the same prediction, identified exactly — so the comparison is between what was said and what happened.
6. Nothing is rewritten. A later result never edits the earlier prediction, and a prediction is never re-created after the fact. If a record is wrong, it stays visible as what it was.

• Settled. The player took part and the official result is known. A player who took part and recorded zero counts as a real zero.
• Did not play. The evidence shows the player did not take part. The prediction stays in the record but is not graded — a missed game is never scored as a zero.
• Unresolved. The evidence cannot yet establish whether the player took part. It stays ungraded until it can.
• Source unavailable. The evidence needed to settle has not arrived yet. The prediction waits; it is not dropped.

Every prospective prediction stays in the record whatever happens — large misses, games where a player recorded zero, players who sat out, and predictions no sportsbook ever quoted. Nothing is removed because it looks bad.

FDI applies this discipline where it tracks a model prospectively. It is a property of how those records are kept, not a claim that any model will perform well: a frozen record can show a model doing poorly just as plainly as doing well, which is the point.