NHL Performance

How the official forecasts have actually done.

264 graded games

2026-27 trial season · frozen model · research only — not betting advice Details

2026-27 SHADOW EVALUATION · prospective · research only.

Forecasts come from the frozen Tournament V2 poisson_glm model. It is not promoted and is not an NHL production system.
This is not a proven betting system and no profit is implied.
Moneyline: replicated historically in Tournament V2 (two protected seasons) and now under prospective shadow evaluation.
Puck line / total: derived from the same score grid; NOT historically validated to the moneyline's standard.
Confidence = probability the predicted outcome occurs (never edge or value).
Starting goalies, where shown, are for information only — the forecast does not use goalies.

Last shadow pass: Sat Nov 07 03:01 PM ET

Quick Read

The moneyline pick was right in 59.5% of 264 graded games.

Average confidence on those picks: 58.0%.

On the same 249 games, the model's log loss was 0.654 against the market's 0.645 (lower is better).

Moneyline

Games	Right	Log loss	Brier	Market log loss (same games)
264	59.5%	0.6573	0.2327	0.6452 (n=249)

By confidence (moneyline)

Confidence	n	Right	Log loss
50.0%–55.0%	105	51.4%	0.692
55.0%–60.0%	73	54.8%	0.690
60.0%–65.0%	44	70.5%	0.619
65.0%–70.0%	30	73.3%	0.585
70.0%–100.0%	12THIN n=12	83.3%	0.472

Puck line and total at the market line

Derived from the same model; not validated to the moneyline's standard.

Market	W	L	P	Win rate ex push
Puck line	153	96	0	61.4%
Total	120	129	0	48.2%

By confidence (puck line and total, at the market line)

Derived from the same model; not validated to the moneyline's standard.

Puck line

Confidence	n	Right	Log loss
50.0%–55.0%	16THIN n=16	43.8%	0.691
55.0%–60.0%	32	53.1%	0.706
60.0%–65.0%	39	59.0%	0.675
65.0%–70.0%	59	61.0%	0.678
70.0%–100.0%	103	68.0%	0.638

Total

Confidence	n	Right	Log loss
0.0%–50.0%	10THIN n=10	40.0%	0.716
50.0%–55.0%	165	49.7%	0.698
55.0%–60.0%	63	49.2%	0.705
60.0%–65.0%	11THIN n=11	27.3%	0.817

ⓘ How these are measured

PROSPECTIVE 2026-27 SHADOW record. Separate from, and never pooled with, the historical Tournament V2 validation (development; Test #1 CONFIRMED; Test #2 REPLICATED). No model changes because of these numbers.

market = multiplicative de-vig of the consensus moneyline in the official snapshot

THIN marks a figure on fewer than 30 graded games (decided games, for the puck line and total).

Puck line and total tables use the pick at the market line only. Confidence there is the probability the pick covers; push mass is not included in it. Pushes are left out of Right and Log loss (n counts decided games) and shown in a P column when any exist. Log loss scores the pick against win/loss using the confidence conditioned on no push, confidence ÷ (1 − push probability), which equals the stated confidence at half-point lines.