College hockey betting angles backlog (2026-09-29)

Context. The model loses to the market (FINDINGS_MODEL_VS_MARKET.md): log losslog lossA score for probability forecasts that punishes confident wrong answers harshly. Lower is better. 0.625 vs 0.603. Fading it also loses. Anything that squeezes more prediction out of the same stats is dead. These angles are timing ones, or use information the line may not have.

Market data. OddsPapi historical odds are free and archived from 2026-01. DraftKings is on ~55% of D-I games; Pinnacle, BetRivers and Caesars on fewer. Snapshots are sparse; the last pre-game one is a median ~6 h before puck drop.

1. Early-season roster turnover ⭐ chosen first

• Idea: October lines lean on last season, but college rosters turn over heavily. The portal is ~4× as costly per point as graduation (backtests/player_movement/FINDINGS_TRANSFERS.md).
• Signal: returning production (points and goalie minutes kept), split by how it was lost (portal vs graduation vs pro), plus incoming transfers' production.
• Test: backtest 2021-2025 against a last-season baseline (no odds exist for past Octobers). Then a forward test against October-November 2026 lines, with thresholds fixed in advance.

2. Friday/Saturday goalie rotation

• Data: goalie_game_stats covers ~5,200 games, 2020-2025. ⚠ Its started column is 0 on every row (never populated): take the starter as the goalie with the most minutes in the game. ⚠ Exclude the EMPTY NET rows.
• Idea: some teams reliably start their backup on Saturday. Does a line posted on Friday price that in? Measure the backup-vs-starter save% gap, and how predictable the rotation is.
• Catch: we can't see starters ahead of time, so this is only a timing bet if the rotation is predictable from history.

Result (2026-09-30, goalie_rotation.py, 2020-2025, no odds): - Rotation is predictable: teams that rotated in ≥50% of earlier weekends rotate 51% of the time, others 24%. - Actual rotation costs: −5.8 win-probability points vs pre-game EloElo ratingA rating system, originally from chess, that moves a team up or down based on results and the strength of the opponent. (CI −9.1 to −2.6), and +0.4 goals allowed. - The predictable version is weak: "rotators'" second games run −1.8 pp vs Elo (CI −5.1 to +1.5), others +1.3. That ~3 pp gap is about the book's cut, measured against Elo rather than the market, and Saturday starters are often known before the line settles. - Verdict: a lead, not an edge. Revisit only if we can see lines before starters are announced.

3. Book-vs-book price gaps (no model)

• Idea: bet a DraftKings or BetRivers price when it beats Pinnacle's no-vig price by more than the book's margin. This is the line-shopping approach that has worked elsewhere.
• Data: Pinnacle priced only 51 games last season. Needs the 2026-27 weekly backfill to judge.

4. Travel and schedule spots

• Examples: Alaska trips, Thursday/Saturday series, the second game after long travel, rest differences.
• Test: five seasons of results WITHOUT odds first. Only if an effect survives out of sample, check it against lines.
• Prior: probably already priced. The Retrosheet split miner found 96.6% of apparent patterns died on new data.
• Result (2026-09-30, schedule_spots.py) — CLOSED.
• The four factors were fixed in advance: rest difference, visitor at Alaska, first game after Alaska, 10+ day break.
• Fit on 2020-23 and tested on 2024-25, they made forecasts WORSE: +0.0027 log loss (CI +0.0009 to +0.0045).
• Alaska looked large in training, but there were only 30 test games; not cherry-picked after the fact.

Rejected

• Regulation-tie (3-way) market. Books price a tie at ~21% with the margin removed; last season 18% of games went to overtime. The three-way margin is 9-12%. Per-book results were noise (±15-20 pp).
• More prediction from the same stats (better xG, save% regression, etc.). That's the approach that just lost.