The 2024-25 Stanley Cup Final ended with Florida beating Edmonton in six games to repeat as champions. The playoffs were a grind — high-seed teams held serve in the West, mid-seed teams ran the East bracket, and the Cup Final came down to two teams the model had rated essentially equal.
I wanted to know what our NHL win-probability model would have produced if we had run it on every playoff game before puck drop, using only data available prior to Round 1. No look-ahead. No retroactive feature work. Just the model we shipped, applied to every game of last year's 86-game postseason.
This post is the backtest. Every round, the Stanley Cup Final breakdown, and a clean read on what the model did well and where it struggled.
The Headline
52 correct out of 86 games. 60.5% accuracy.
Broken out by round:
- East 1st Round (21 games): 12/21 (57.1%)
- West 1st Round (26 games): 20/26 (76.9%)
- East 2nd Round (12 games): 7/12 (58.3%)
- West 2nd Round (11 games): 6/11 (54.5%)
- East Conference Final (5 games): 1/5 (20.0%)
- West Conference Final (5 games): 3/5 (60.0%)
- Stanley Cup Final (6 games): 3/6 (50.0%)
For reference, here are public benchmarks for NHL playoff accuracy:
| Source | Typical playoff accuracy |
|---|---|
| MoneyPuck playoff model | 58-62% |
| FiveThirtyEight NHL Elo | 55-60% |
| Pinnacle closing-line favorite | 60-64% |
| Chalk (always pick higher seed) | 56-60% |
| Coin flip | 50% |
Our 60.5% is squarely in the MoneyPuck/Elo zone. NHL is the highest-variance major sport in the playoffs — best-of-seven series are frequent underdog outcomes, and goaltending randomness swings series. A 60% accuracy number is what you'd expect a serious calibrated model to produce.
The Standout: West 1st Round 20/26 (77%)
The model's best round of the entire postseason.
The West bracket was the more predictable conference: Winnipeg, Dallas, Edmonton, and Vegas entered with strong ELO ratings and played like it. Our pre-game win probabilities on those matchups were consistently in the 60-70% range, and the actual results matched — higher-ELO teams closed out their series convincingly.
The model's view going in was "the West is better and more predictable this year." That view was correct.
Where the Model Got Outclassed
East Conference Final: 1 of 5 (20%) — the worst single round of the postseason.
This series featured two teams that were close on regular-season ELO but with very different playoff performances. Our model had the higher-seeded team at ~60% to win each game; they lost four of the five games they played. That's a classic situation where the model's pre-game priors didn't reflect a real but hard-to-quantify factor: the higher seed had accumulated injuries late in the regular season that carried into the playoffs.
1/5 is unlucky on a small sample, but it's consistent with what we already flagged as a weakness in the NHL model: it doesn't currently ingest injury data, and the playoffs magnify the cost of missing that.
The Stanley Cup Final: 3 of 6 Correct
Florida beat Edmonton 4-2 in the Final. Our per-game predictions:
| Game | Location | Model P(home) | Actual | ✓/✗ |
|---|---|---|---|---|
| Game 1 | @ EDM | 55.0% | EDM 4-3 FLA | ✓ |
| Game 2 | @ EDM | 55.0% | FLA 5-4 EDM | ✗ |
| Game 3 | @ FLA | 55.0% | FLA 6-1 EDM | ✓ |
| Game 4 | @ FLA | 55.0% | EDM 5-4 FLA | ✗ |
| Game 5 | @ EDM | 55.0% | FLA 5-2 EDM | ✗ |
| Game 6 | @ FLA | 55.0% | FLA 5-1 EDM | ✓ |
The model had the Final as a coin flip. Not because it was indecisive — because it was right. Florida and Edmonton were within 5 ELO points of each other entering the series (both in the 1560s), so the model correctly assigned ~55% home probability (just HFA) to whichever team hosted.
Over six games, the model went 3-3, which is exactly what a coin-flip prediction should produce. Calibration without discrimination is the term of art: the model knew it didn't know who would win, and said so.
The series-level prediction was more useful: Florida had a slight ELO edge (1566 vs 1562), which translated to ~52% to win the series. Florida won 4-2, so the model's series-level lean was correct, just thin.
What the Model Got Right Structurally
Three wins worth flagging beyond raw accuracy:
-
West 1st Round at 77%. Four of the eight series were essentially called, and the three that went the "wrong" way were all close to coin-flips in the model's view. When the model has strong signal, it was right. When it had weak signal, it was honest about it.
-
Series-level vs game-level predictions. The Cup Final coin-flip is a case in point. The model correctly identified that a series between two evenly-matched teams would go ~6-7 games and could tilt either way. Public fan polls had this at 65%/35% in Edmonton's favor because of their higher offensive profile. Our model was more honest.
-
Home-ice advantage sized correctly. NHL HFA is smaller than NBA (~40 ELO points vs 80). Over 86 games, the model's home team pre-game win probability was tightly clustered around 53-58%, which matches the historical home win rate in the playoffs almost exactly.
What the Model Needs to Improve
Three specific things the backtest exposed:
- Goaltender-specific features. The single biggest miss category. Our current model uses team-level save percentage; we need individual starting-goalie stats that update between rounds.
- Injury-adjusted priors. The East Conference Final blew up because a key player was playing through injury. Injuries matter 2-3x more in playoff hockey than in the regular season because of the compressed schedule. We don't currently feed injury data at all.
- Momentum-cascade modeling. In NHL series, after a 2-1 deficit the trailing team's win rate drops sharply because of cumulative fatigue and emotional load. Our model currently treats each game independently — we should consider a series-state component.
The Takeaway
The model called the West correctly, coin-flipped the Final honestly, and missed the East Conference Final on injury-related factors it doesn't model. Overall 60.5% is competitive; the training ECE of 6.59% is what we'd expect a calibrated model to produce.
Next up: our live 2026 Stanley Cup futures — published today, updated as Round 1 progresses. If you want to backtest your own NHL strategies against the same snapshot data, you can pull live NHL edges via the API — 7-day free trial, no credit card.
Related Reading
- 2026 Stanley Cup Futures (Live) — Every 2026 playoff team's championship probability, updated weekly.
- How to Build an NHL Prediction Model — The NHL-tuned ELO + special-teams features pipeline behind this backtest.
- NBA Finals 2025 Retrospective — The NBA equivalent (84 games, 59.5%).
- World Series 2025 Retrospective — The MLB equivalent (47 games, 59.6%).
- Calibration Beats Accuracy — Why our Cup Final coin-flip prediction was calibrated even when "wrong".
Data sources: ESPN NHL game data (public); ELO computed from game results with K=8 and HFA=40 (NHL-tuned). All 86 postseason games were held out of the ELO training set. Pre-game predictions use the deployed wp_model_NHL.pkl. The full prediction table is reproducible from the /v1/backtest endpoint.