New to these reports? Start here
- Dotted-underlined words have a plain-English definition — hover or tap them. Every term is also on the glossary page.
- "Null" means we found nothing, not that something broke. Most reports here are negative results, on purpose — knowing an idea doesn't work is the point.
- Two questions get asked separately. First, is the effect real? Second, is it already priced into the betting odds? An effect can be completely real and still useless to bet on.
- A "calibration" row is a self-check. It runs the same method on something already known to be true. If that fails, the whole report is unreliable — so it's reported alongside the findings.
- If a confidence interval includes zero, the real effect might be nothing at all, so no claim gets made.
Moneyline accuracy by series position (2026-10-17)
Script: series_position.py → series_position_results.json. Series = the same two teams meeting
twice within 2 days; same rink = same home team both nights, home-and-home = home team swaps.
Regular season only (March+ conference playoffs reported separately in the JSON). Ties dropped.
The DB's neutral_site/playoff_game flags are all 0, so neutral sites can't be split out.
Walk-forward model (2022-23, 2023-24, 2025-26 folds; out of sample)
| Group | n | Accuracy | Log losslog lossA score for probability forecasts that punishes confident wrong answers harshly. Lower is better. |
|---|---|---|---|
| Same rink, game 1 | 851 | 60.5% | 0.658 |
| Same rink, game 2 | 833 | 62.9% | 0.647 |
| Home-and-home, game 1 | 169 | 63.3% | 0.652 |
| Home-and-home, game 2 | 173 | 66.5% | 0.655 |
| Single games | 679 | 65.2% | 0.631 |
| All regular season | 2,708 | 63.0% | 0.647 |
Game 2 minus game 1: same rink +2.4pp (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. −2.4 to +7.0); home-and-home +3.2pp (−6.8 to +12.5). No position or series type is reliably better or worse.
Deployed model (2025-26 only, stored pre-game predictions): same rink g1 56.6% (n=106) vs g2 71.4% (n=105), CI +2.5 to +27pp — one season, ~100 games a cell, one of several cuts; the walk-forwardwalk-forwardEvaluating week by week using only what was knowable before each week, mimicking how the model would actually have been used at the time. sample is 8× larger and shows +2.4. Treat as noise. Note 73 of its 94 same-rink game-2 predictions were made before game 1 was played (Tue/Sat cron).
Game-1 winner in game 2 (walk-forward, same rink)
Model gave the game-1 winner 54.2% in game 2; they won 58.2% (n=780, +4.0pp, SE 1.8). By season +5.6 / +0.7 / +5.5pp. Home-and-home: −0.8pp (n=156, SE 4.0). Part of it is a general home biasbiasWhether the misses lean consistently one way. A projection can have a good average error size but still be biased if it is almost always too high. Bias is often the more fixable problem.: the model averages 51.9% for the home team, home teams win 55.8% (regular season); game-1 winners are home 57% of the time. Net of that, ~+3.5pp remains whether the game-1 winner is home or away. EloElo ratingA rating system, originally from chess, that moves a team up or down based on results and the strength of the opponent. already updates after game 1, so the remaining signal would be same-weekend information (goalie, lineup, matchup). Found in a post-hoc cut: test forward on 2026-27 before using it. Picking the game-1 winner (58.2%) is still worse than the model (62.7%).