Methodology

overall = clamp( 50 + 15 · composite_z , 0 , 99 )

1. Aggregate the player's weekly regular-season rows into season totals and rates.
2. Shrink each rate toward the cohort mean by sample size: x_adj = (n·x + k·μ) / (n + k). This stops a backup QB with 12 great throws from topping the list. Volume stats aren't shrunk.
3. z-score across the cohort. Stats where lower is better (INTs, fumbles) are negated.
4. Weighted-average the z-scores. A component whose stat doesn't exist for a player (e.g. CPOE before 2006) is dropped and the rest renormalized.
5. Re-standardize that weighted average within its own cohort (mean 0, standard deviation 1) before mapping it onto the 0–99 scale. A weighted average of several already-clipped z-scores doesn't itself have a standard deviation of exactly 1, and how far it varies from 1 differs by position — so without this step, positions with fewer or more correlated components would quietly get a different effective scale than positions with more independent ones. This step is what lets one global formula reach both ends of 0–99 fairly for every position.

Box-score stats catch splash plays but miss coverage, gap discipline, and snap share. Treat DEF as 'how much did this player show up in the box score', pending Phase 2 participation data. All metrics per game.

• Defense is box-score only.
• No opponent adjustment — a strong performance vs. a weak defense looks the same as vs. an elite one.
• No aging curve, no multi-year prior — each season stands alone.
• Pre-2006 receivers lose the CPOE / air-yards components.
• No team-defense or special-teams rating yet.

A standard NFL Elo model (FiveThirtyEight style): K-factor 20, a 55-point home-field bonus, the margin-of-victory multiplier with the autocorrelation fix, and a 40% between-season regression to 1500. Win probability is the logistic function of the Elo gap; projected margin is the gap divided by 25.

v1.5 adds a starting-QB overlay. Each QB carries a rolling, walk-forward value (an EPA/dropback average, shrunk toward a replacement level on his first tracked start). When a team's actual starter differs from the QB who's been starting for them, that game's probability and margin shift by the value gap between them, scaled to Elo points — calibrated by backtesting every game since 1999. It's zero whenever the normal starter plays, so most games are unaffected; the team's underlying Elo rating itself is never changed by it.

Across ~7,300 games since 1999 it picks the straight-up winner ~64% of the time (Brier ~0.22) — in line with published NFL Elo models. ~66–67% is roughly the ceiling for any pregame model. The QB overlay's gain is modest and shows up mainly in calibration (Brier/log-loss), not the straight-up rate — see the backtest numbers on the track-record page. See the full track record →

A Monte-Carlo simulation of the season, ~20,000 runs. Each run plays out every remaining game (margin drawn from a Normal around the Elo projection, SD ~16), seeds each conference, and plays the bracket. Aggregating gives each team's odds to make the playoffs, win its division, get the #1 seed, and reach / win the Super Bowl.

Simplified tiebreakers (win% → head-to-head → division record → point diff); Elo is static through the sim; injuries and trades aren't modelled.

1. Recent form — the player's per-game average last season.
2. Regress toward a position baseline by sample size (heavy prior — 6–8 games).
3. Adjust for what the opponent allowed to that position, clamped to ±30%.
4. Distribution — Poisson for counts (receptions, TDs), Normal for yardage. Output is a mean, the 10th–90th percentile range, and P(stat > line).

Preseason projections lean entirely on last year — no new-team fits, rookies, or injuries yet. Players with under 4 games last season are skipped; no OL, defense or kicker projections.

Win probabilities and projections are informational estimates with real uncertainty — not advice for betting, daily fantasy, or any decision where an error costs money. See the Terms of Service.