What we knew after week 21

SRS is a ridge-regularized simultaneous solve: every team’s rating is its average scoring margin adjusted for the ratings of everyone it played, with margins capped so a single blowout cannot dominate a season. Elo walks the schedule forward one game at a time from a fixed 1999 epoch. Between seasons it keeps 55% of each team’s distance from average and regresses the rest back toward the middle — last season is evidence, not a starting point. They are independent systems over the same games, published side by side because they disagree in informative ways.

Every board is cut through its own week — it absorbs the games played by that week and nothing later. That is what makes an old week worth reading: it is what was knowable at the time, not a retrospective re-rating. The hindsight companion shows how far these ratings sat from what the finished season says was true.

These do not beat the market. On the held-out 2018-2025 seasons — 2,227 games the constants were never tuned on — Elo predicts game margin at 10.236 MAE and SRS at 10.295 (over its 2,099 games; SRS produces nothing in week 1), against 9.829 for the closing line. Read them as a description of what teams have done, not as a price.

Two families of column in the underlying artifact are deliberately not published here: a normalized srs_z, which divides by the season’s final spread and so would let a week-5 z-score know how the season ended, and the per-season era descriptives, which are season-final by construction. Neither can honestly appear on a board labelled “after week 21.”