Context is usually the biggest single input
Most model inputs change slowly. Availability does not. A starting pitcher scratched, a quarterback ruled out, a guard's minutes restriction, a fighter withdrawn and replaced on short notice — each can change the relevant probabilities more than weeks of form data.
| Change | What moves |
|---|---|
| Starter out (NFL/NBA/WNBA) | Usage, targets and minutes redistribute to teammates; team totals move |
| Lineup position (MLB) | Plate appearances: a leadoff hitter gets more trips than a No. 9 hitter |
| Goaltender change (NHL) | Save-rate expectations and game totals |
| Short-notice replacement (UFC) | The whole matchup: a different opponent, often a different weight or camp |
| Rounds rescheduled (UFC) | Distance and finish probabilities (3 vs 5 rounds) |
"Reported" is not "confirmed"
Card changes are shown only from verified sources. A reported withdrawal alone keeps a bout active with a warning; a bout is removed only on confirmed evidence. When a specific reason is not in the verified sources, the page says so — "A specific reason is not recorded in our verified sources" — rather than guessing. Injury pages show one card per availability episode, not one per headline, with no inferred diagnosis.
The same discipline is worth copying: act on confirmed information, treat rumours as uncertainty, and never let a model silently assume a fighter or player is in when the evidence says otherwise.
Timing and the closing line
Markets absorb confirmed news fast. The value of an availability read is mostly in when you know it relative to the market — which is why Closing line value is the natural way to measure whether you processed news well.
Late scratches can happen after a model has locked. Official-record policies differ by product — some void the call, some grade it — and each product page states its rule.