Altitude

What it measures

Altitude is our answer to "how did he actually play this week," built position by position from what a single game can honestly support. Every weekly stat carries some mix of skill and luck, and we measured the mix before choosing the ingredients: halve a player's snaps at random, compute the stat on each half, and see whether the two halves agree. Stats that cannot agree with themselves inside one game do not get to grade anyone here.

For quarterbacks, Altitude is EPA per dropback with the least reliable outcomes priced out. A single game of interceptions is about one percent signal, so interception plays are re-priced at that season's mean incompletion value; a sack fumble's bounce is a coin flip, so those plays are priced as clean sacks. For receivers, tight ends, and backs it is earned yards: targets priced at the league rate for their depth, carries at the league rate for their box count. Opportunity earned is the player; weekly conversion is mostly context and luck, so conversion is priced, never observed.

The published number is a historical percentile from 0 to 100 against every qualifying game at the position since 2016, with fixed bands: Grounded below 10, Turbulence below 40, Level below 60, Cruising below 90, Soaring at 90 and above.

How it is computed here

Each position has its own view (qb_altitude_week, rb_altitude_week, wr_altitude_week, te_altitude_week) with the raw inputs published beside the score, so the adjustment is always visible. The QB view carries both raw and de-noised EPA per dropback; the receiver views carry earned yards next to delivered yards, and the gap between them is its own diagnostic. Qualifying gates are stated in each view: 15 or more dropbacks for quarterbacks, 10 or more routes for receivers and tight ends, 8 or more touches for backs, regular season only.

The design rule throughout: delete provable luck, never add identity. An early composite version scored higher on reliability by leaning on rushing identity, and it flunked what we call the absurdity test by ranking an MVP season 22nd at the position. Reliability bought with identity is not a game grade, so we shipped the plain version.

How fast it stabilizes

Measured, not guessed. Single-game reliability is 0.147 for the QB score against 0.099 for passer rating and 0.109 for raw EPA per dropback; 0.449 for TE, the highest in the family; 0.356 for WR against 0.267 for raw yards; 0.307 for RB against 0.201 for delivered scrimmage yards. The QB score reaches half signal in about six games. Even the best weekly number is mostly weather: treat one week as a report on that week, and let the season tell you who the player is.

How it gets misused

• Reading adjacent bands as a ranking. Cruising versus Level is often noise; the band is the claim, and the gap between neighboring games is usually not.
• Treating receiver Altitude as a catch grade. The declared referent is opportunity earned, not outcome delivered. A receiver who earns elite depth-priced targets from a backup quarterback can out-score one whose targets an MVP converted; that is the referent working.
• Expecting it to grade what the data cannot see. Pass protection assignments, untargeted coverage, and pre-snap gravity are not in play-by-play, so they are not in Altitude. Naming the blind spots is part of the metric.