PuckCast Edge
Every pick is free. The Edge is the rest.
Betting edges, stake sizing, props priced against the line, fantasy tiers and the graded ledger. $14.99 a month in season, or $99 for the year.
PuckCast Edge
Betting edges, stake sizing, props priced against the line, fantasy tiers and the graded ledger. $14.99 a month in season, or $99 for the year.
v2.0 en95+trend
The complete record for every PuckCast projection since the model went live. Win probabilities shown against implied market odds. We don't hide the losses.
Certified, season by season
2021-22 onward, each season a full 82-game schedule. The shortened 56-game 2020-21 season is in the certified headline above but left out of this table so no row here is a partial season next to five full ones.
| Season | Games | Accuracy | Log loss | Brier |
|---|---|---|---|---|
| 2025-26 | 1,312 | 55.4% | 0.6816 | 0.2443 |
| 2024-25 | 1,312 | 59.9% | 0.6613 | 0.2347 |
| 2023-24 | 1,312 | 61.2% | 0.6557 | 0.2320 |
| 2022-23 | 1,312 | 61.4% | 0.6527 | 0.2305 |
| 2021-22 | 1,312 | 64.7% | 0.6463 | 0.2271 |
Accuracy by Grade
+0.14σ
Headline Z (30d)
61.2%
Last 30 days · 201g
60.7%
Baseline (walk-forward)
+0.5pp
Delta vs baseline
Recent accuracy is within ±1.5σ of the walk-forward baseline. Normal noise.
This is a 30-game-a-month view, which is a short window and a noisy one. Model health runs the same check over whole seasons, where the picture is worse: the edge over simply picking the home team has fallen from +11.1pp at its peak to +3.2pp last season.
Calibration
Live accuracy per confidence grade for the 235 games published this season, against the walk-forward baseline the grade was trained against.
The model uses 3 site grades (A/B/C). Backend grades (A+ through F) are mapped to these 3 tiers. Small-sample grades (<5 games) are hidden because their accuracy isn't informative yet.
Seasonality
Aggregated from 235 graded predictions across the season archive. The model warms up in early season, cools mid-winter, and tightens up into the playoffs.
Sep
50%
4/8
Oct
61%
133/218
Nov
56%
5/9
Notable picks
We show losses too. The most-confident hits, the most-confident misses, and the closest calls of the 235-game season. Every season, filterable →
Latest Results
5 of 9 correct over the last 2 game days (Nov 1, 2026 – Nov 2, 2026).
| Matchup | Result | Pick | Grade | Outcome |
|---|---|---|---|---|
| VGK@BOS | 2-3OT | BOS50.7% | C | HIT |
| DET@FLA | 5-3 | FLA56.3% | B | MISS |
| COL@CHI | 5-2 | COL67.3% | B | HIT |
| CGY@SJS | 2-1 | CGY50.7% | C | HIT |
| WPG@CAR | 0-7 | CAR75.6% | A | HIT |
| MIN@WSH | 5-1 | MIN53.2% | C | HIT |
| UTA@PHI | 3-6 | UTA64.5% | B | MISS |
| EDM@NYR | 4-1 | NYR56.2% | B | MISS |
| OTT@ANA | 3-6 | OTT50.7% | C | MISS |
By Team
Live-season accuracy on each team's games. Hot teams are where the model's read has been most reliable; cold teams are where you should weight picks lightly.
Computed over 32 teams with 5+ graded games this season. See each team's page for the same number in context, or filter the full results history →
Understanding the Numbers
Plain-English explanations for the stats on this page.
Accuracy
The percentage of games where our predicted winner actually won. The NHL baseline (always picking home) is ~53.7%. We beat it by +7.0 points.
Log Loss
Measures calibration — when we say "70% chance", it should happen ~70% of the time. Lower is better. 0 = perfect, 0.693 = random guessing.
Edge (points)
How far from 50/50 each prediction is. A 65% pick has 15 pts of edge. Higher edge = higher confidence. Our grading system is based on edge size.
Brier Score
Mean squared error between predicted probability and actual outcome. Combines accuracy and confidence into one number. 0 = perfect, 0.25 = random.
Baseline
The accuracy from always picking the home team. Any useful model should beat this. Home teams win ~54% in the NHL.
Holdout Testing
Testing on games the model never saw. For each season, the model trained only on prior years. True forward validation that prevents overfitting.
Common Questions
Overall accuracy is 60.70% across 7,428 walk-forward tested games. A-grade picks (highest confidence) hit at 75.7%.
A testing method where the model is trained on past seasons and tested on future games it has never seen. This prevents overfitting and gives realistic accuracy estimates.
Live results update daily as games complete. The track record page shows every prediction ever made.
Accuracy is the percentage of games where PuckCast's predicted winner actually won. It's tested across 7,428 holdout games using 6-fold walk-forward validation, where the model only sees past data when making each prediction.
Log loss measures calibration — when PuckCast says 70%, those teams should win about 70% of the time. Lower log loss means more honest probability estimates. A perfect model scores 0, random guessing scores 0.693.
Holdout results are from testing v2.0 en95+trend on 7,428 historical games the model never saw during training. Live results are actual predictions made this season before each game. Both use the same model -- live just validates it in real time.
Live results update daily as games complete. Holdout test results are from v2.0 en95+trend and remain fixed as the model's validation benchmark.
