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.
Certified record: 60.7% over 7,428 games (2020-21 to 2025-26). Every pick is published before puck drop and graded after, wins and losses both. Here's what the number means and how to check it yourself.
TL;DR
Certified record: 60.7% over 7,428 games (2020-21 to 2025-26). Hockey is a low-ceiling sport to predict: a coin flip gets 50%, home ice alone gets around 53.7%, and Vegas closing lines run 58-60%. Every pick PuckCast makes is published before the game and graded after, and the misses stay on the public ledger right alongside the hits.
Every NHL game has two possible winners, so the floor for any prediction is 50%. That makes accuracy a deceptively small-looking number in hockey compared to, say, picking winners in a sport with heavy favorites. The useful comparison isn't 50% to 100%; it's 50% to the realistic ceiling.
Home ice advantage alone is worth about 53.7% accuracy with zero modeling. Vegas closing lines, priced by people with money on the line, land around 58-60%. Public hockey models that disclose their validation method report 55-65%. PuckCast's 60.70% sits inside that range, measured the honest way: on games the model never saw during training.
4,509 correct calls out of 7,428 graded regular-season games, spanning 6 seasons (2020-21 to 2025-26), including the shortened, 56-game 2020-21 COVID season. That window is reproduced from a committed parquet file in CI on every run (scripts/verify_headline.py), so the number on this page cannot drift from what actually shipped.
It isn't one flat number either. Confidence grades split it out: A-grade picks (the model's clearest edge) hit 75.7%, B-grade hits 61.4%, and C-grade, the marginal calls, hits 53.4%, close to a coin flip. Averaging them together into one headline hides that spread. See what 61% accuracy means for the full breakdown.
One more distinction worth being precise about: 60.70% is the certified historical result, not a forecast for the season ahead. It won't move again until this window is re-certified. The honest forward-looking number, at a typical season's level of parity, runs lower. Model health carries both figures side by side and says why they differ, rather than only showing the flattering one.
A pick is generated and published before that game's puck drop, using only the information available at that moment: rosters, rest, injuries, form, and the rest of the model's 98 features. Nothing about the pick changes once the game starts.
After the final score, the pick is marked correct or incorrect against the actual result and written to the public ledger. That's the entire grading rule: no adjustment, no exclusion, no re-grading with hindsight.
What's in the certified window
Because a track record that only shows wins isn't a track record. 2,919 of the 7,428 graded games in the certified window were misses, and every one of them is still in the ledger, next to the date, the teams, and the probability the site displayed before the game. /results highlights the most-confident misses right alongside the most-confident hits, on purpose.
The model was wrong on 2,919 games it graded itself. Publishing that number is the whole point of grading in public instead of only pointing at the wins.
Nothing on this page asks you to take a headline on faith. The full, filterable ledger is public at /results/ledger: every graded game, sortable by season, grade, team, and result, including the misses. The certified 60.70% figure is reproduced from a committed parquet file every time the site's tests run, not typed in by hand anywhere. And /model-health tracks calibration and whether the edge is decaying, updated in the open rather than only when it's flattering.
If you want the mechanics of how a pick becomes a probability in the first place, how the model works covers the features, the validation method, and why a simpler model beat more complex ones in testing.
PuckCast is the only public NHL model that publishes every pick before puck drop and grades every one of them in the open: 60.70% across 7,428 games, 6 seasons (2020-21 to 2025-26). "Most accurate" is hard to verify for a model that doesn't show its misses. This one does, at /results/ledger.
It means PuckCast picked the winning side 4,509 times out of 7,428 graded games. A coin flip gets 50%. Home ice alone gets roughly 53.7%. Vegas closing lines run 58-60%. Hockey has a low ceiling because goals are rare and a single bounce can flip a result, so 60.70% sits well ahead of a naive baseline without promising anything close to certainty on any single game.
Each pick is generated and published before that game's puck drop, using only information available at that moment. After the final score is in, the pick is marked correct or incorrect and added to the public ledger. Nothing is regraded, removed, or reworded after the fact.
Yes. The full ledger at /results/ledger lists all 7,428 graded games, wins and losses both, filterable by season, grade, team and result. /results highlights the most-confident hits and the most-confident misses side by side.
No. It's the certified result of 6-fold walk-forward validation across 2020-21 to 2025-26, and it doesn't move until that window is re-certified. The honest forward-looking number, at a typical season's parity, is lower; /model-health carries both and explains why they differ.