The first football board of the season tempts offseason models outside the markets they were built to predict. Margin models belong on margins, totals models on combined scoring, and player models on player outcomes. Confidence should not widen the job.
Build the Week 1 watchlist around market fit instead. Give every model one job, one timing boundary, and one live status. A model that has no bet is allowed to be correct.
Start with the predicted outcome
State what the model predicts in plain language instead of the algorithm, input list, or nickname.
- Spread: scoring margin, cover probability, or another team-versus-team outcome.
- Total: combined scoring or another game-level scoring outcome.
- Prop: a player-level opportunity or performance outcome.
If the predicted outcome cannot be stated cleanly, the model is not ready for the watchlist. “General football strength” is not a market.
Shared inputs do not make outcomes interchangeable
Pace can matter to spreads and totals. Quarterback quality can matter to both, and team context can inform a prop. None of that means one predicted outcome has been validated for every market touched by the same inputs.
A spread model tries to separate opponents, while a totals model describes the shared scoring environment and a prop model connects team context to one player’s role and outcome. Those are different outcomes with different scoring rules.
Reuse raw ingredients when useful. Rebuild and validate each predicted outcome separately.
Put every model in a live lane
Native-market ready
The predicted outcome matches the market, every input exists before the wager time, the backtest is point-in-time safe, and the live scoring path matches the tested path. Ready means eligible for observation or a controlled deployment under the bankroll rule; the label does not establish an edge.
Research only
The idea is plausible but one boundary is unresolved: a feed changed, an input timestamp is unclear, or a live market differs from the historical definition. Keep producing predictions if that helps diagnosis, but do not let the model borrow authority from incomplete validation.
Wrong market
The model answers a different question. A margin model used on a receiving prop belongs here until a player-level predicted outcome and evaluation exist. This label is useful because it keeps a vague signal from becoming a confident bet.
Check timing before performance
A correct predicted outcome can still be unusable live when an input arrives too late. Record when every input becomes available, not merely whether the final dataset contains it.
Postgame efficiency can train future games. It cannot describe the game that created it. A revised injury designation can help after publication but was unavailable at the decision time. A closing price can grade an earlier wager but cannot be smuggled into its prediction.
“Present in the table” and “available to the bettor” are different states.
Watch for information that arrives too late
A model that looks brilliant on spreads, totals, and props can be seeing part of the answer rather than a breakthrough.
Inspect inputs finalized after the event, rolling fields that include the predicted-outcome row, end-of-season summaries joined backward, and market information captured after the intended wager time. Remove the suspicious input and rerun the test.
A result that fails under that challenge stays in research. The backtest did its job.
Use Studio to expose the job
Rebuild or inspect the workflow in Studio so the predicted outcome, inputs, market mapping, and grading are visible together. Grade a margin model on margin, a totals model on the total, and a player model on the player outcome represented in its historical rows.
Studio leads into the same model-building flow when you want to inspect an existing idea. A visible predicted outcome is better than a memorable nickname because nicknames survive after definitions drift.
Separate insight from authorization
A model can provide useful context without issuing a bet. A spread model can show team separation while the totals model remains uncertain, and a player model can identify role growth while the game model sees no side.
Keep those outputs separate. When the event is graded, review each model against its own predicted outcome. A blended recommendation makes every model look responsible when it wins and none responsible when it loses.
The Receipts Drawer
- Predicted outcome: the exact result the model predicts.
- Native market: the market that matches the predicted outcome.
- Decision time: when the prediction is intended to be used.
- Point-in-time inputs: information available before that boundary.
- Backtest scope: historical window, training split, evaluation split, and scoring rule.
- Known failure modes: role changes, sparse history, feed gaps, or scheme changes.
- Live status: native-market ready, research only, or wrong market.
- Pause condition: the observation that moves the model out of deployment.
The card should identify the model version. When the predicted outcome or preprocessing changes, create a new version instead of editing history.
Review process before record
After Week 1, ask whether the model ran on time, whether inputs were complete, whether the market matched, and whether the prediction distribution resembled the backtest. Those questions can reveal a deployment failure immediately.
Then record the graded result under its window and sample. Do not promote a model because one wager won or demote it because one lost. An opening result is one event, not a market license.
Keep the mapping stable enough to learn
If a quiet spread signal sends the model hunting for a total, the season log measures improvisation. Keep the native market fixed until a deliberate new outcome definition is built and evaluated.
Shared inputs are fine. Shared credibility is not automatic. A new market means a new model card, even when much of the pipeline is reused.
Bottom line
A useful Week 1 watchlist gives every model one job it can perform. Match spread models to spreads, total models to totals, and player models to player outcomes; verify timing and label research honestly. Preserve a wrong-market state instead of forcing a pick.
When the board opens, discipline means letting a model say nothing outside its lane. That silence is cheaper than a confident answer to the wrong question.
NFL ATS cover-margin distribution
Bars count completed NFL schedule rows by closing-spread cover margin using the repository canonical home-margin grading convention.
