NHL Prediction Model Performance & Calibration

See how our NHL prediction model performs: accuracy, Brier scores, calibration, and error metrics across thousands of games.

To understand each metric, read Understanding Performance Metrics . To apply this in practice, view today's NHL predictions, our NHL playoff odds, and in-game win probability charts.

How to Read These Metrics

Accuracy: The percentage of games where the predicted winner (team with >50% win probability) actually won. Simple but incomplete — it ignores how confident the model was.
Brier Score: Measures the mean squared error of probability predictions (0–1 scale, lower is better). A coin-flip baseline yields 0.25; our model targets values below 0.24. Brier score rewards well-calibrated confidence levels, not just picking the right side.
Log Loss: A logarithmic scoring rule that heavily penalises confident wrong predictions. Assigning 90% to a team that loses costs far more than assigning 55%. This keeps the model honest about uncertainty.
Calibration: Shows whether stated probabilities match real outcomes. In a well-calibrated model, games given a 70% win probability should be won about 70% of the time. The calibration tables below group predictions into decile bins so you can verify this directly.
MAE / RMSE (Total Goals Error): Both measure how far predicted totals are from actual totals, in goal units. The game-level windows report MAE (mean absolute error) because the point estimate is the median of the simulated total, and the median minimizes MAE. The cross-validation folds report RMSE because they score the mean (expected goals), where RMSE is the consistent metric.

Why calibration matters most: For probabilistic predictions, calibration is more important than raw accuracy. A model that says "55%" every game can be 55% accurate but useless for decision-making. A well-calibrated model tells you how much to trust each prediction. Learn more in our analytics guide and methodology.

Game Predictions (Multi-Window)

Window	Start	End	Games	Accuracy	Brier	Log Loss	Avg Winner Prob	MAE Total
last 30	2026-03-17	2026-04-16	245	56.7%	0.2352	0.6616	53.0%	1.893
season to date	2025-10-01	2026-04-16	1312	61.8%	0.2275	0.6464	53.6%	1.834
multi season	2023-10-10	2026-04-16	3936	61.5%	0.2288	0.6488	53.5%	1.859

Totals (Over 5.5)

Window	Games	Accuracy	Brier	Log Loss	Avg Outcome Prob
last 30	245	57.6%	0.2458	0.6854	52.1%
season to date	1312	57.2%	0.2462	0.6860	51.7%
multi season	3936	56.0%	0.2485	0.6904	51.2%

Playoff Game Performance

Start	End	Games	Accuracy	Brier	Log Loss
2024-04-20	2026-06-14	256	60.9%	0.2348	0.6619

Current Matchups

NJD vs PHI model odds NYR vs TBL model odds CBJ vs BUF model odds DET vs NYR model odds CAR vs WSH model odds WPG vs BOS model odds

Daily Performance

Show:

Date	Games	Accuracy	Brier	Log Loss

Prediction Recap Highlights

Definition: High confidence means the model assigned the predicted team a win probability well above 50%. Edges that hit are the highest-confidence correct calls, misses are the highest-confidence incorrect calls, and surprise results show the largest absolute gap between win probability and the actual outcome.

Show:

Biggest Model Edges That Hit

Date	Matchup	Win Prob	Outcome

Biggest Misses (High Confidence)

Date	Matchup	Win Prob	Outcome

Surprise Results

Date	Matchup	Surprise	Outcome

Calibration (Win Prob Deciles) — Last 30

Bin	Count	Mean Pred	Observed
2	4	28.0%	25.0%
3	13	36.1%	30.8%
4	73	45.7%	46.6%
5	74	54.8%	45.9%
6	56	64.0%	60.7%
7	20	74.9%	80.0%
8	5	82.6%	80.0%

Calibration (Win Prob Deciles) — Season To Date

Bin	Count	Mean Pred	Observed
2	17	26.8%	5.9%
3	76	36.5%	21.1%
4	351	45.8%	40.5%
5	462	54.8%	51.3%
6	285	64.0%	67.4%
7	107	73.8%	80.4%
8	14	81.9%	78.6%

Calibration (Win Prob Deciles) — Multi Season

Bin	Count	Mean Pred	Observed
2	49	26.3%	12.2%
3	240	36.2%	28.3%
4	1017	45.7%	42.8%
5	1437	55.0%	53.8%
6	851	64.2%	67.7%
7	302	73.8%	79.1%
8	40	82.1%	90.0%

Calibration (Over 5.5) — Last 30

Bin	Count	Mean Pred	Observed
5	65	57.6%	55.4%
6	168	63.6%	58.9%
7	12	71.3%	50.0%

Calibration (Over 5.5) — Season To Date

Bin	Count	Mean Pred	Observed
4	9	47.6%	55.6%
5	587	57.0%	54.9%
6	689	63.2%	59.7%
7	27	71.6%	51.9%

Calibration (Over 5.5) — Multi Season

Bin	Count	Mean Pred	Observed
4	25	48.1%	56.0%
5	2238	56.7%	55.3%
6	1632	62.8%	57.4%
7	41	71.6%	51.2%

Team Calibration (Home, Top 15 by Volume)

Team	Count	Mean Pred	Observed	Bias
EDM	148	62.4%	62.2%	+0.2%
CAR	147	69.2%	70.7%	-1.5%
DAL	146	55.3%	63.0%	-7.7%
FLA	146	57.4%	62.3%	-5.0%
VGK	143	60.0%	60.1%	-0.2%
COL	138	66.6%	66.7%	-0.1%
WPG	133	54.2%	62.4%	-8.2%
BOS	133	51.7%	56.4%	-4.7%
TOR	133	51.3%	54.1%	-2.9%
MTL	133	51.6%	49.6%	+1.9%
TBL	132	59.4%	62.1%	-2.7%
WSH	131	54.5%	58.8%	-4.3%
NYR	131	50.9%	51.9%	-1.0%
MIN	131	52.8%	52.7%	+0.1%
VAN	130	49.3%	43.1%	+6.2%

Team Calibration (Pred vs Observed)

Mean Pred Observed

Starter Calibration (Home)

Window	Starter Status	Games	Accuracy	Brier	Log Loss
last 30	Starter	245	56.7%	0.2352	0.6616
season to date	Starter	1312	61.8%	0.2275	0.6464
multi season	Unknown	18	61.1%	0.2328	0.6580
multi season	Starter	3918	61.5%	0.2288	0.6488

Cross-Validation (Expanding Window)

Summary: 3 folds | Brier: 0.2549 | Log Loss: 0.7033 | RMSE Total: 2.393

Show fold details

Fold	Train N	Val N	Brier	Log Loss	RMSE
Fold 1	703	2,089	0.2561	0.7059	2.431
Fold 2	1,396	1,396	0.2550	0.7036	2.369
Fold 3	2,094	698	0.2534	0.7003	2.380

In-Game Checkpoints — Last 30

Checkpoint	Games	Accuracy	Brier	Log Loss
end_p1	17	64.7%	0.2108	0.6250
end_p2	17	76.5%	0.1587	0.4810
ot_start	5	60.0%	0.2178	0.6143
p3_10	17	94.1%	0.0727	0.2360
p3_5	17	88.2%	0.0824	0.2520
pregame	17	41.2%	0.2552	0.7059

In-Game Checkpoints — Season To Date

Checkpoint	Games	Accuracy	Brier	Log Loss
end_p1	1394	65.8%	0.2111	0.6083
end_p2	1394	77.7%	0.1483	0.4510
ot_start	348	63.8%	0.1956	0.5583
p3_10	1394	84.0%	0.1033	0.3231
p3_5	1394	85.9%	0.0880	0.2766
pregame	1394	56.0%	0.2429	0.6786

In-Game Calibration — Pregame (Last 30 Days)

Bin	Count	Mean Pred	Observed
4	6	44.8%	50.0%
5	6	54.1%	16.7%
6	2	63.5%	50.0%
7	3	73.5%	66.7%

In-Game Calibration — End P2 (Last 30 Days)

Bin	Count	Mean Pred	Observed
0	4	8.4%	0.0%
1	2	11.1%	50.0%
2	1	28.8%	0.0%
3	1	35.6%	0.0%
4	2	44.0%	50.0%
5	1	52.2%	0.0%
6	1	64.0%	100.0%
7	1	78.8%	100.0%
8	2	83.8%	50.0%
9	2	98.4%	100.0%

In-Game Calibration — P3 10 (Last 30 Days)

Bin	Count	Mean Pred	Observed
0	5	2.6%	0.0%
1	3	16.6%	0.0%
2	1	25.5%	100.0%
3	1	37.5%	0.0%
4	1	48.4%	0.0%
5	1	55.6%	100.0%
8	1	89.1%	100.0%
9	4	96.7%	100.0%

xG Holdout — Contextual

Train: 2023-10-10 – 2025-12-27 | Test: 2025-12-28 – 2026-06-14

Games (test): 794 | Shots (test): 68188 | ROC AUC: 0.785 | Log Loss: 0.2211 | Brier: 0.0602

xG Splits — Contextual Strength State

Split	Shots	Goal Rate	AUC	Log Loss	Brier
Even	54578	6.3%	0.780	0.2022	0.0541
PP	11485	10.6%	0.723	0.2915	0.0807
PK	1491	7.2%	0.838	0.2132	0.0609
EmptyNet	634	50.3%	0.754	0.5979	0.2059

xG Splits — Contextual Shot Type

Split	Shots	Goal Rate	AUC	Log Loss	Brier
wrist	28569	7.2%	0.816	0.2060	0.0561
snap	17624	8.6%	0.774	0.2516	0.0711
slap	8178	4.8%	0.720	0.1788	0.0443
tip-in	6597	6.4%	0.666	0.2270	0.0582
backhand	5044	8.6%	0.818	0.2325	0.0644
deflected	1103	11.5%	0.702	0.3273	0.0951
wrap-around	407	5.4%	0.751	0.1820	0.0456
bat	357	7.8%	0.779	0.2346	0.0625
poke	217	8.8%	0.693	0.2704	0.0701
between-legs	48	12.5%	0.810	0.2950	0.0832
nan	37	59.5%	0.686	0.8924	0.2869
cradle	7	14.3%	1.000	0.2298	0.0636

xG Holdout — Neutral

Train: 2023-10-10 – 2025-12-27 | Test: 2025-12-28 – 2026-06-14

Games (test): 794 | Shots (test): 68188 | ROC AUC: 0.782 | Log Loss: 0.2246 | Brier: 0.0615

xG Splits — Neutral Strength State

Split	Shots	Goal Rate	AUC	Log Loss	Brier
Even	54578	6.3%	0.779	0.2026	0.0543
PP	11485	10.6%	0.703	0.3101	0.0881
PK	1491	7.2%	0.836	0.2131	0.0607
EmptyNet	634	50.3%	0.747	0.5975	0.2069

xG Splits — Neutral Shot Type

Split	Shots	Goal Rate	AUC	Log Loss	Brier
wrist	28569	7.2%	0.811	0.2099	0.0576
snap	17624	8.6%	0.770	0.2556	0.0726
slap	8178	4.8%	0.720	0.1789	0.0444
tip-in	6597	6.4%	0.661	0.2277	0.0584
backhand	5044	8.6%	0.804	0.2419	0.0680
deflected	1103	11.5%	0.705	0.3258	0.0946
wrap-around	407	5.4%	0.737	0.1868	0.0469
bat	357	7.8%	0.774	0.2388	0.0640
poke	217	8.8%	0.704	0.2674	0.0693
between-legs	48	12.5%	0.794	0.3017	0.0866
nan	37	59.5%	0.694	0.8424	0.2857
cradle	7	14.3%	1.000	0.2357	0.0685

Monthly Performance Trends

Track how model performance varies month-to-month across the season.

Month	Games	Accuracy	Brier	Log Loss
2023-10	140	57.1%	0.2383	0.6678
2023-11	213	56.8%	0.2422	0.6767
2023-12	219	61.6%	0.2394	0.6717
2024-01	208	54.3%	0.2358	0.6619
2024-02	172	61.0%	0.2369	0.6664
2024-03	228	62.3%	0.2267	0.6455
2024-04	132	53.0%	0.2495	0.6910
2024-10	166	67.5%	0.2128	0.6145
2024-11	220	63.6%	0.2213	0.6328
2024-12	214	65.9%	0.2097	0.6090
2025-01	224	63.4%	0.2346	0.6618
2025-02	122	56.6%	0.2259	0.6412
2025-03	234	65.4%	0.2169	0.6238
2025-04	132	65.9%	0.2308	0.6537
2025-10	180	66.7%	0.2098	0.6088
2025-11	225	63.1%	0.2210	0.6320
2025-12	226	65.5%	0.2255	0.6422
2026-01	240	60.8%	0.2284	0.6481
2026-02	74	67.6%	0.2255	0.6455
2026-03	242	52.9%	0.2505	0.6947
2026-04	125	61.6%	0.2236	0.6376

Playoff Model Performance

Game-level and series-level accuracy across playoff rounds.

Playoff Games

Round	Games	Accuracy	Brier	Log Loss
All Rounds	82	56.1%	0.2455	0.6841
Round 1	45	57.8%	0.2456	0.6843
Round 2	22	59.1%	0.2417	0.6761
Round 3	9	44.4%	0.2568	0.7069
Round 4	6	50.0%	0.2426	0.6781

Playoff Series

Round	Series	Accuracy	Brier	Log Loss
All Rounds	15	66.7%	0.2298	0.6520
Round 1	8	62.5%	0.2352	0.6636
Round 2	4	75.0%	0.2320	0.6567
Round 3	2	50.0%	0.2213	0.6313
Round 4	1	100.0%	0.1950	0.5827

Playoff Calibration (Pred vs Observed)

Mean Pred Observed