Équipes
Registre des prédictions
Chaque prix d'avant-match enregistré avant la mise en jeu initiale, chaîné par hachage et noté le lendemain.
Public ledger · built 2026-10-18
Chaque prédiction, inscrite avant la mise en jeu, notée après.
Le prix du matin et, quand les formations sont publiées, le prix à l'étape des formations, pour chaque match, ajoutés à un fichier chaîné par hachage avant le match et notés le lendemain matin. Rien ici ne peut être modifié après coup sans briser la chaîne.
- Notés
- Justes
- Dossiers
- Tête de chaîne
Résultats en direct
Par niveau de confiance
| Niveau | Détails |
|---|---|
| coin flip | |
| lean | |
| strong | |
| solid | |
| heavy |
By game type: regular season ; preseason .
Les plus récents d'abord
games recorded
| Date | Match | Prix du matin (domicile) | Prix à l'étape des formations | Clôture du marché (domicile, sans marge) | Niveau | Choix | Final | Juste ? |
|---|---|---|---|---|---|---|---|---|
| 2026-10-18 | CBJ at MIN | 62% | strong | MIN | — | |||
| 2026-10-18 | CAR at CGY | Non disponible | strong | CAR | — | |||
| 2026-10-17 | NJD at WSH | 53% | lean | NJD | no | |||
| 2026-10-17 | CAR at WPG | 35% | strong | CAR | yes | |||
| 2026-10-17 | VAN at TBL | 76% | heavy | TBL | yes | |||
| 2026-10-17 | SEA at FLA | 63% | strong | FLA | yes | |||
| 2026-10-17 | NYI at TOR | 54% | coin flip | TOR | no | |||
| 2026-10-17 | BUF at MTL | 56% | coin flip | MTL | yes | |||
| 2026-10-17 | SJS at DET | 57% | solid | DET | yes | |||
| 2026-10-17 | OTT at PHI | 42% | solid | OTT | no | |||
| 2026-10-17 | NYR at CBJ | 55% | coin flip | CBJ | yes | |||
| 2026-10-17 | DAL at CHI | 34% | heavy | DAL | yes | |||
| 2026-10-17 | STL at NSH | 50% | coin flip | STL | no | |||
| 2026-10-17 | BOS at LAK | 56% | solid | LAK | no | |||
| 2026-10-17 | EDM at UTA | Non disponible | lean | UTA | no | |||
| 2026-10-17 | PIT at VGK | Non disponible | strong | VGK | yes | |||
| 2026-10-16 | PIT at COL | 65% | solid | COL | yes | |||
| 2026-10-16 | BOS at ANA | Non disponible | coin flip | BOS | yes | |||
| 2026-10-15 | BUF at TOR | 52% | lean | BUF | yes | |||
| 2026-10-15 | NYI at OTT | 60% | solid | OTT | yes |
Market baseline
Model against market
The market is the hardest baseline there is: the vig-free closing line, the last price before puck drop, median across books. Same games, same rule: log loss on who wins (overtime and shootout included). Lower is better.
| Comparison | Détails |
|---|---|
| model against market |
57 games carry a closing line from 47 market snapshots. Vig removed by proportional (multiplicative) normalisation per book: p_home = (1/d_home) / (1/d_home + 1/d_away); consensus = median across books. Closing line: the last market snapshot before puck drop; quotes updated after puck drop are dropped. Rule, registered before any line was stored (p4-9-market-logloss-v1): Reported every night whatever the sign: model log loss, market log loss and the per-game difference with a game-cluster bootstrap 95% interval (2,000 resamples, fixed seed). Wording: below 100 paired games "too few games to call"; from 100 games "model beats market" only if the interval lies below zero, "market beats model" only if it lies above zero, otherwise "not separable". Nothing is adopted or tuned from this comparison.
Every family, proper scores
Model against a dumb baseline
Lower is better. Each family is scored on its proper rule (log loss for yes/no calls, ranked probability score for ordered outcomes) against marcel-style league rates: home-win share, regulation goals and finish shares of the three seasons before the forecast season, weighted 5/4/3 by games. Ranges are 95% confidence sequences, valid on any night you read them. Samples are games.
| Family | Détails |
|---|---|
| win | |
| finish | |
| score | |
| total | |
| prop-goal | |
| prop-point |
Picks: 90 of 144 right (Wilson 95%: 0.544 to 0.700). 13816 standing assertions, 8 on games not yet final, 2876 void under their rule (a skater who did not dress). Results are re-checked against the official boxscore at 72 hours and 14 days (121 and 48 so far, 1 revisions). Season simulation: no season-simulation forecast has been frozen in the ledger yet; the scorers (crpsNormal, crpsEnsemble, midPit) are in place and tested.
Committed slate
What was promised, and what was missed
Each morning the list of games and families to be forecast is committed to its own chain before any forecast; every later freeze carries the slate's hash. A game family with no morning freeze before puck drop scores as the baseline here, so a skipped night cannot flatter the record.
| Date | Détails |
|---|---|
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-05 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-07 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 | |
| 2026-10-13 |
Player props
Points props by decile and star tier
| Decile | Détails |
|---|---|
| 1 | |
| 2 | |
| 3 | |
| 4 | |
| 5 | |
| 6 | |
| 7 | |
| 8 |
By star tier (rank by prior points per game among the club's dressed skaters): star (top 2): model 0.645 against baseline 0.660 over 144 games; top six: model 0.685 against baseline 0.691 over 144 games; depth: model 0.554 against baseline 0.553 over 144 games.
Challengers in shadow
Is anything beating the champion?
Every register challenger is priced on the same games as the champion. Each one's likelihood ratio is an e-value; e-BH controls false alarms across all of them, every night. An alarm raises a decision card; nothing is promoted automatically.
| Challenger | Détails |
|---|---|
| ratings-only | |
| isotonic-calibrator |
No alarm (e-BH threshold 60; champion calibration e-value 0.212; backtest band 0.626 to 0.681). Starting goalies known for 134 of 150 games (1 named before puck drop).
Bias audit
Does the price lean anywhere?
Every slice of games (club, home and away, favourite, tier, back-to-back, month, division, season): the model's average probability against the share that actually won, with bootstrap ranges, false-discovery control across slices, and an anytime-valid sequential monitor that keeps its error rate under daily peeking.
Backtest, games: slices, flagged at 95% ( survive false-discovery control): month 02 (home side) . Full table on /integrity.
Calibration
Do 60% calls win 60% of the time?
- perSeason
- pooled
The tier record above is the live version of the same question; the model card on /models carries the six-season version.
Provable by someone other than us
External timestamps on the chain heads
Each day the two chain heads are hashed into one root and timestamped by an RFC 3161 authority. Anyone can verify that today's predictions existed before today's games without trusting this site.
- 2026-10-18 verified · freetsa · serial · Oct 18 12: · root
- 2026-10-17 verified · freetsa · serial · Oct 17 12: · root
- 2026-10-16 verified · freetsa · serial · Oct 16 13: · root
- 2026-10-15 verified · freetsa · serial · Oct 15 14: · root
- 2026-10-14 verified · freetsa · serial · Oct 14 13: · root
- 2026-10-12 verified · freetsa · serial · Oct 12 12: · root
- 2026-10-11 verified · freetsa · serial · Oct 11 12: · root
Vérification
Freeze audit (2026-10-18): records are covered by a public timestamp issued before their game started; earlier records were anchored once a day and some receipts came after puck drop. Since 2026-10-07 every freeze is timestamped as it is written.
Every pre-game price is appended to an append-only file before puck drop, each record carrying the SHA-256 of the previous one; results are appended the next morning the same way. tools/ledger.mjs verify recomputes both chains. Chain heads at build time: predictions ( records), results ( records). The model card is on the measures page; the daily email carries the same prices.