Read the interval, not just the number.

Each projection is three numbers. The prediction is the model's point estimate of PPR points for that week. The floor and ceiling are the point estimate shifted by the lower and upper residual quantiles observed on a validation season for that position. A wide interval means the model has historically been less precise for players like this one; a narrow one means the opposite.

Half-PPR and standard totals are derived on the server by subtracting a share of the player's estimated receptions, so switching the scoring control never changes the underlying model output.

Tiers group players whose prior-season profile looks alike. They are fitted once before the season and do not update week to week. The probability next to each player is the mixture's confidence that they belong to that tier rather than a neighbouring one; a low value means the player sits near a boundary. Tier order follows mean prior-season points per game, so a lower tier number is the stronger group.

The evaluation page compares the champion to a baseline that only knows a player's own earlier points. The gap between the two columns is the honest measure of what the model adds. Positional cohorts show where the model is stronger or weaker; the rolling-origin folds show how error moves through a season when the model is refit before every scored week.

How accurate are the projections?

Exactly as accurate as the evaluation page says for the cohort you care about. No accuracy figure is written into this site by hand; every number is read from the committed evaluation artifact.

Should I always start the higher prediction?

Not necessarily. If you need a safe result, a higher floor can matter more than a higher point estimate. The start/sit tool shows both so you can decide.

How often do predictions change?

A scheduled job runs once per week during the season. It publishes a new predictions file only when its policy allows it, and the manifest records what it decided and why.

Does the model know about injuries, weather, or news?

No. The only input is nflverse weekly data. Missed games show up in a player’s history after the fact, but there is no injury report or weather feed.