QB+ is a quarterback projection metric scaled so that 100 always equals an average NFL starter — the same idea as wRC+ in baseball. A QB+ of 120 means the quarterback projects 20% better than the median starter; 80 means 20% worse. Because it's re-scaled each season, you can compare across years.

Inspiration: These projections owe a debt to Tom Tango's Marcel system — the famously simple baseball forecasting method that proves you don't need complexity to be accurate. Marcel uses three years of history, regresses toward the mean based on playing time, and applies an age adjustment. That's the whole thing. It remains one of the hardest baselines to beat in sports forecasting, precisely because it doesn't try to be clever. We built a Marcel for football first. Beating it required considerably more machinery than we expected — and the margin is narrower than our egos would prefer.

How this model works: The projection engine follows Marcel's philosophy — recent performance, regression, aging — but replaces hand-tuned weights with a Bayesian hierarchical model, a technique widely advocated by statistician Andrew Gelman, trained on hundreds of quarterback-seasons of historical data. The model learns how much to trust each input signal, and because it's hierarchical, quarterbacks with limited track records are naturally pulled toward the league mean while established starters are projected primarily from their own history.

EPA is a team metric. The model treats EPA/play as a team-level outcome — the product of quarterback skill, coaching, and supporting cast — and projects it with coaching effects baked in. That's the field-level number you'd actually see. But QB+ needs to measure the quarterback alone, independent of his environment. To isolate individual skill, we project five QB-isolated metrics — completion percentage over expected (CPOE), quantified film grade, air EPA, deep completion rate, and pressure-to-sack rate — and map them into EPA-equivalent units via a cross-sectional model. Projected metrics are regressed toward league average based on sample size (plays and career seasons), so quarterbacks with shorter track records receive more conservative skill estimates. The gap between total projected EPA and the QB-isolated component is the coaching effect: how much a quarterback's environment inflates or deflates his raw numbers. Think of it like park factors in baseball.

Credible intervals: Click any player name to see their historical trend and an 80% credible interval on their projection — meaning 80% of model-consistent outcomes fall within the shaded range. Quarterbacks with shorter track records get wider intervals (more uncertainty). Established starters get narrower ones.

What you're looking at: Every row is a pre-season projection — what the model forecast heading into that season, based only on data available beforehand.