Four things the residual audit surfaced that don't go away under adjudication. Each one is a real signal in the per-pick residuals, not a leak the model should fix, but a piece of the football landscape the data is pointing at.
The GM tenures where the published verdict most disagrees with the record alone. Rookie classes the model refuses to bury, one-great-afternoon legends it refuses to canonize, and in every case the same machinery: partial pooling declining to let a handful of picks prove anything.
Read →The aggregate "small-school picks beat expectation" story doesn't actually hold once you correct the residual baseline. What does hold: certain GMs find dramatically more talent at small schools than others. Jerry Jones is the league outlier; Larry Allen out of Sonoma State is the marquee.
Read →Pro Bowl and All-Pro residuals correlate with the team a player is drafted into, even after stripping out per-pick performance. Top of the table: BAL, KC, DAL, SF, PIT. Bottom: JAX, NYJ, CIN. The pattern looks more like franchise quality than market size.
Read →Plot career-starts residuals by 5-year buckets and the passing era shows up in the floor: TE and WR residuals climb in the 2020s, RBs stay negative across forty years, OL flattens after a 1990s peak. The model is allowed to surface this: the s(draft_year) smoother absorbs league-wide era shocks but not position-specific drift.
Read →Two trends survived adjudication for a quieter reason: they're what the model was designed to capture. "Never-played picks" cluster at the bottom of y_cgs because that's the signal. OFF/DEF/ST residuals differ slightly because ST has 217 picks across 45 years. Neither is a bug.
Read →