Pre-registration: roster turnover vs the October market (2026-27)

Frozen 2026-09-29, before the season's first puck drop on 2026-10-02. The commit containing this file, forward_2026_team_ratings.json and roster_turnover_forward.py is the record. Nothing in them changes after 2026-10-02. Any later change goes in an "Amendments" section below, dated, with both results reported.

Question

Early-season lines lean on last season. Does a team's roster turnover carry information those lines don't already have? Turnover here means: - scoring returning and scoring lost to the portal; - scoring brought in by transfers; - goalie minutes returning; - NHL-drafted freshmen.

Backtest that motivated it (roster_turnover.py, no odds)

• Data: 2022-2025, Oct-Nov games, 1,299 games, ratings frozen preseason, fit leave-one-season-out.
• Headline: adding turnover to a regressed-Elo prior improved log losslog lossA score for probability forecasts that punishes confident wrong answers harshly. Lower is better. by 0.0033 (95% CI95% confidence intervalThe range the true value is plausibly in. If this range includes zero, we cannot rule out that the real effect is nothing at all. −0.0138 to +0.0072). It helped in 3 of 4 seasons.
• Weights (EloElo ratingA rating system, originally from chess, that moves a team up or down based on results and the strength of the opponent. points per 1 SD): returning scoring +21, portal losses −21, incoming transfers +17, goalie minutes +5, drafted freshmen +46.
• Two choices made after seeing the first run, both disclosed:
• 2021 dropped: its prior season was COVID-shortened, and the ratios exploded.
• Features clipped at the 1st/99th percentile.
• Two bugs found and fixed before these numbers:
• The goalie feature was empty: goalie_game_stats.started is never populated, so minutes are used instead.
• An over-strong ridge penalty pinned every weight to about 1 point.

Inputs, frozen

• Team ratings: forward_2026_team_ratings.json holds each team's regressed preseason Elo and turnover shift. It uses rosters scraped from CHN on 2026-09-29 and coefficients fit on 2022-2025.
• Odds: OddsPapi snapshots from the Monday cron (oddspapi_backfill.py --recent-days 10). Book preference: Pinnacle > DraftKings > BetRivers > FanDuel > Caesars.
• Open = first stored pre-game snapshot; last = last one. Both are de-vigged proportionally.
• Games: 2026-10-02 to 2026-11-30, D-I vs D-I, both teams in the frozen file, a pre-game line present, ties dropped.

Tests, decided now

1. PRIMARY — does turnover add to the market? - Build p_adj = logistic(logit(p_last) + 0.5 × shift × ln10/400), with shift = home turnover shift − away turnover shift. The 0.5 is an a-priori guess that the market has already priced half. - Pass: the paired-bootstrap 95% CI on mean log-loss difference (p_adj − p_last) lies entirely below 0.
2. SECONDARY — closing-line value. - Take games where the model's probability differs from the de-vigged opener by ≥ 3 pp and the line has more than one snapshot. Take the model's side. - CLVclosing line valueWhether you got a better price than the market settled at. Widely used as a faster signal of skill than profit, which takes ages to measure. = how far the market moved toward that side from open to last, in de-vigged probability. - Pass: the 95% CI on mean CLV lies entirely above 0. - ROIreturn on investmentProfit as a percentage of the money wagered. +2% means $2 profit per $100 bet. at the opening price is reported but is NOT a pass criterion: it is too noisy at this sample size.

Decision rule

• Graded once, after the 2026-12-07 cron has captured November's games. Interim looks are allowed; changes are not.
• Both fail: the angle is closed.
• Either passes: extend the same frozen rule through the season before any money is considered. One pass at ~150-250 games is a lead, not an edge.
• Expected sample: ~150-250 games. Last season OddsPapi had a line for ~30% of D-I games.

Amendments

(none)