Skip to main content
ProbetricsProbetrics
NHL track record · Model

How the NHL model is tested

The method itself isn't published here. What is published: every pick is locked before puck drop and graded after the final (overtime and shootouts count like any other result), the model is evaluated on seasons it never saw, and a change only ships if it holds up there. Winners are the record: the pick is the side the model favors, with the probability it gives that side. When sportsbooks price a game, that probability follows the betting market's price: in testing, the market beat the team ratings alone, so the ratings decide only games the books haven't priced yet.

Backtest

Walk-forward: every game priced from ratings that have seen only the games before it, fixed before these seasons were ever scored. This is the ratings alone; anchored to the market's price, as picks are published now, held-out games 2022–26 went 60.0%.

Backtest winners · ratings aloneEvery game priced only from what was known before it, then scored once. Kept apart from the record.

Walk-forward backtest of the ratings alone, held-out seasons 2019–26: not in the record
Won–lost
4,987–3,523
Win rateWhere the true win rate likely sits (90% range). Under 30 picks, too early to read.
58.6%
90% range 57.7–59.5%
Graded
8,510
picks
Win rate 58.6%90% range

For scale: always picking the home team would have won 53.6% of the same games. Hockey games sit closer to a coin flip than basketball or football, so most calls land between 50% and 65%.

Calibration

When the model gives a side 60%, that side should win about 60% of the time. The lanes put the live picks (only a handful so far: NHL picks are just starting to grade) beside the backtest for each band.

When it said 70%, did it win 70%?Each lane is a confidence band. Calibrated means the dot lands in its gray band; the bar shows luck's range.

live and backtest
  • Said 50–55%
    47% won 7 of 1590% range 28–67%backtest 51.0% of 2,884
  • Said 55–60%
    36% won 4 of 1190% range 18–61%backtest 57.3% of 2,455
  • Said 60–65%
    63% won 5 of 8only 8 graded: too few to judgebacktest 63.8% of 1,621
  • Said 65–70%
    100% won 6 of 6only 6 graded: too few to judgebacktest 66% of 996
  • Said 70–100%
    67% won 2 of 3only 3 graded: too few to judgebacktest 75% of 554
The band it statedLive win rate90% rangeBacktest win rate

Live picks, locked before the start. Backtest: walk-forward backtest of the ratings alone, held-out seasons 2019–26, not in the record. Lanes with fewer than 10 graded picks are faded.

Live record

Winners calledOnly picks locked before the start count. A pick made later, or priced after the fact, never enters it.

Locked before the start, graded after
Won–lost
24–19
Win rateWhere the true win rate likely sits (90% range). Under 30 picks, too early to read.
56%
90% range 43–68%
Graded
43
picks
Expected winsThe sum of its stated win probabilities. If those are right, wins land in this range 9 times in 10.
25.2
90% range 19.9–30.4
Win rate 56%90% rangeIts probabilities said 59%

Won 24. Its stated probabilities expected 25.2 (90% range 19.9–30.4): inside that range, in line with its probabilities.

Betting favorite on the same 41 priced games: 23–18 (ours 23–18).

Against the betting favoriteSame games, winners only. The favorite is the side with the shorter no-vig moneyline when the pick was locked.

41 priced games
Our pickon the games that had a price23–18 56%n 41
The betting favoritethe same games, no model23–18 56%n 41
When we picked the favorite23–18 56%n 41
When we took the other sidethe games we disagreed with the market0–0 —none

Every pick here is from the market-anchored era: a priced game's win % is the market's no-vig price. 2 picks have no stored price and sit in none of these rows. Under 30 games a record is too early to read either way. Every pick, one row each →

Every graded pick, the ledger and expected vs actual winners →

What we don't publish

The method, its parameters, and what was tried and rejected are not published — that's the part of the work we keep to ourselves. What's always public is the result: every pick locked pre-game, graded after, calibration shown honestly, and nothing about the record edited after the fact.

For information only, not betting advice. 21+ only. Gambling problem? Call or text 1-800-GAMBLER.