How it works

Every call for week W is made using only data available before week W kicked off. Trailing form and usage features are computed over prior games and explicitly exclude the week being predicted; the game environment comes from pre-kickoff market lines. The historical track record is a walk-forward backtest: predictions for each season are made by a model trained only on earlier seasons, exactly as it would have run live.

That means the hit rates you see are honest out-of-sample results, not a model graded on games it already saw. When you browse a past week, you're seeing what the model would have said before those games were played.

Players are ranked within their position by predicted PPR points for the week. The rank maps to a call: the top band is 🟢 START, the next band is 🟡 FLEX (startable in deeper lineups or as a flex), and the rest is 🔴 SIT. Players ruled out are forced to SIT regardless of rank.

The confidence number shown next to a call is not a model guess; it is the real historical rate at which that call bucket produced a startable finish, from the walk-forward backtest:

Read it plainly: a START call has finished startable 57% of the time; a SIT call only 12%. The gap between those two numbers is the product. No call is a guarantee: 57% means roughly 43% of START calls miss.

A player's week is startable if they finish at or above the positional startable line in PPR scoring: top 12 for QB and TE, top 24 for RB and WR (one starter, or two, per position in a 12-team league).

A call hits when it was right about that line: a START/FLEX player finishes startable, or a SIT player doesn't. The overall hit rate across all 15,216 graded calls is 77%, but be skeptical of that headline number: most calls are SITs, and SIT calls are usually right. The calibration table above is the honest view, which is why it leads every page.

One deliberately simple model per position: a ridge regression with learned (not hand-set) weights over point-in-time features, trained walk-forward so each season is predicted using only earlier seasons. The features:

• Recent form: trailing PPR over the last few games
• Target share: trailing share of team targets
• Opportunities: trailing carries + targets per game
• Snap share: trailing share of offensive snaps
• Game environment: the team's implied point total from pre-kickoff market lines
• Defense vs. position: trailing PPR the opposing defense allows to the position

Every trailing feature is shifted so the week being predicted is never part of its own inputs. Players with too little history get no call rather than a made-up one.

The site has two kinds of numbers, and we keep them separate on purpose:

• Predictive: the Start/Sit calls. These claim an edge, so they carry receipts: every call is graded and the full track record is public.
• Descriptive: the Stat Explorer leaderboards. These describe what already happened (usage, efficiency, expected-TD gaps). They are useful context, but a leaderboard rank is not a prediction and we don't grade it like one.

If a number on this site claims to know the future, it has a public hit rate attached. If it doesn't, it's describing the past.

Everything is built from a single source: official NFL box scores, snap counts, play-by-play, betting odds, and injury status. One source keeps the pipeline reproducible and the point-in-time guarantee auditable.

It also sets honest limits. That source has no air-yards feed and no official EPA, so the model does not use them, and metrics we derive ourselves from the play-by-play (like the expected-points cuts in the Stat Explorer) are labeled "our EP/EPA" to make clear they are our reconstruction, not the official figures. When a stat can't be built cleanly from this source, we'd rather not show it than fake it.