Caleb Williams

Receiving

Rushing

Passing

nflverse season totals for 2025, joined on normalised name and position (gsis 00-0039918). These are actuals — nothing here is modelled.

The sources that ranked him disagree by 16.0 picks between highest and lowest (standard deviation 5.9).

Mean and median are computed across the sources that actually ranked him — an unranked source never enters the average, because a sentinel treated as a pick number would drag a real player hundreds of slots. Spread, deviation and agreement are null under two ranked sources rather than 0, since one source cannot agree with itself. The FFT model rank is our own output and is never folded in as if it were an outside source.

The weekly LINES are measured: 2025 game logs, joined on gsis id 00-0039918. Everything derived from them above — per-game, standard deviation, coefficient of variation, boom and bust rates, recent form — is computed by us and carries the FFT model (beta) label. It is not an observed stat.

This whole block is on the PPR scale, including the thresholds — the artifact publishes boom and bust in PPR (QB: boom ≥ 25, bust ≤ 15) and converting only the points would leave the rates measured against a different ruler than the one they were defined on. The hero above is half-PPR because that is what the model ranks on. Two rulers, each labelled, rather than one silently mixed. Rates need at least 4 games to carry weight; recent form is the last 4.

Every field is derived by FFT from openly-licensed nflverse play-by-play, or taken from the free public NFL Next Gen Stats endpoint. No licensed trial data (Sportradar, BALLDONTLIE) is included, which is why this file can be committed and shipped with the build. Red zone: Plays snapped inside the opponent 20-yard line. Red-zone usage measures opportunity; this repo’s own backtest shows it adds little on top of plain volume, so it is retained as context and is deliberately not described as an edge.

fft-projection-2.3.0-beta. The model reads measured production, opportunity and age, and reads no ADP field at any stage — so it can disagree with the market, which is the only reason a disagreement is worth anything. Draft capital was not resolved for this player, which the model treats as unknown rather than as undrafted.

• Identity join: Resolved to gsis 00-0039918 via the 2025 nflverse season file. The weekly log and advanced splits are keyed on that id.
• Market ADP: A bundled snapshot of platform ADP, not a live feed. Shown for comparison and never read by the projection model.
• Headshot: Resolved from exact-name.
• Vitals: Height, weight, college and jersey are read from nflverse and keyed on gsis id 00-0039918 — not on name. Name keying matches players to relatives who share one: "Marvin Harrison Jr." also matches his father's 1996 row, and whichever came first in the file would win.
• Not shown: draft round and pick. The bundled player file carries them on roughly 48% of rows and populates them late, so a blank is equally consistent with a genuine undrafted free agent and with a first-rounder nobody has ingested yet. The model’s own three-state read is used instead, above.