Penalty kill, goalies, shot quality: what actually predicts NHL games

Broadcasts love a penalty-kill percentage or a hot goalie. We tested the popular ideas against our results-only Elo to see which ones add anything.

Research as of Nov 20, 2026.

The short answer. Of everything we tested against our results-only Elo, shot quality (expected goals) is the strongest lead: a typical advantage is worth about 3 points of win chance, but it did not clear our bar on the one forward test season. Goalie quality showed no detectable value. The penalty kill failed outright because the percentage barely repeats. Backup goalies cost about 1.5 points, not clearly different from zero, and an early look at back-to-backs, not yet independently checked, suggests rest matters, with the size still uncertain. And the pattern that runs through all of it: effects measured on past seasons came out smaller, or not detectable at all, on seasons we had not touched.

1. How we tested

Our NHL Elo rates teams from results alone. For each idea we asked a simple question: if we add this to Elo, are the predictions more accurate on games the idea was not fitted on? We wrote the pass rule down first, fitted on four seasons, tested on the fifth, and repeated. A pass needed the held-out season to improve clearly and four of five seasons to go the right way. “Underpowered” means it did not meet the bar but was not a clear fail either. “Fail” means no sign of help.

IdeaVerdict
Shot quality (5v5 expected goals share)Underpowered, best lead
Goalie quality (goals saved above expected)Underpowered, no detectable value
Special teams (power play and penalty kill)Fail
All three combinedUnderpowered, mostly shot quality

2. Shot quality: the strongest lead

Expected goals (xG) scores each shot by where it was taken from and how, so a team that gets better chances wins that battle even when the puck does not go in. Teams that win the 5v5 xG share won slightly more than their Elo said: about 4.4 Elo points per percentage point of share. A one-standard-deviation advantage (4.6 points of share) is worth about 20 Elo points, or about 3 points of win chance.

It improved predictions in four of five seasons. But on the pre-registered held-out season, 2025-26, the gain was small and could be chance, so it did not pass. It is also not independent of Elo: shot share and Elo are correlated at 0.76, which is why adding it on top doubles up some of the same information. And shot share settles quickly, reaching a split-half correlation of 0.5 by about 20 games, whereas goal share needs about 70 games. That is a reason shots can be a steadier signal than goals. We are not adding it to the frozen Elo; the next step is a forward test on the rest of 2026-27.

3. Goalies

Two separate checks. First, backup goalies. A team starting its backup won about 1.5 points less often in a test of three seasons, on the fairest measure, and that estimate’s 95% range includes zero, and its opponents' skaters scored a little more (anytime-goal rate up about 3% in relative terms, about half a percentage point). Shots on goal did not change. Adding goalie information to a win-prediction model made it less accurate, not more.

Second, goalie quality, measured as goals saved above what the shots were worth. Even using the goalie who actually started (not always known in advance), the held-out season got slightly worse, and the pooled effect was zero: no detectable value. This is not proof goalies do not matter; goalie results simply do not repeat much. At about 25 appearances per half-season the two halves agree at a correlation of only 0.24. For more see NHL starting goalies.

4. The penalty kill

Power-play and penalty-kill percentages added nothing: they improved predictions in one of five seasons. The reason is that the penalty kill barely repeats. Even over a full 82-game season the correlation between its odd and even games is about 0.30; the power play is about 0.38.

A concrete case: we were asked to check a claim that Colorado's penalty kill was about 69%. On Nov 19, the league stats had Colorado at 80.9% for 2026-27 (13 goals against on 68 times shorthanded), 11th of 32. Nothing we could find, whether the first 15 games, the last 10, home or road, comes near 69%. And if a team really did go 69% over 55 shorthanded situations, shrinking it toward the league rate gives about 77.1%, against a league average of 78.9%. That is worth roughly 0.05 goals a game. Small samples exaggerate.

5. Back-to-backs: an early result, not yet independently checked

This study has not been independently checked yet. On 2021-26, a team playing the second night of a back-to-back did clearly worse than Elo expected: about 7.7 points of win chance (95% interval 4.9 to 10.4). It improved predictions in all five fit seasons.

But on 2018-19 and 2019-20, seasons we had never used, it did not improve predictions. Teams on a back-to-back did still underperform there, winning 44.0% of games where our Elo said 47.5%, but by about half as much as in the fit seasons (40.7% against 47.9%). Its pre-registered test on those seasons failed, so we have not proven a size: about 7.7 points in the fit seasons, about 3.5 in the untouched ones. In a separate, small comparison with the closing line this season and late last season, the line gave a back-to-back team about 3 to 5 points less than our Elo. For that comparison see why NHL betting lines move and our Elo model doesn't.

6. The pattern: effects shrink on fresh seasons

Across studies the same thing keeps happening. Back-to-backs (an early, not yet independently checked study): about 7.7 points in the fit seasons, about half that on the untouched ones. Star players: sizes fitted on 2021-26 showed no detectable gain on 2018-20 (see how much a missing NHL star is worth). Shot quality: clear gains in 2021-22 and 2022-23, about none in 2023-24 and 2024-25, and a small gain that could be chance in the 2025-26 forward test. The lesson is not that nothing matters. It is that a number fitted on a past stretch of NHL games usually overstates the next stretch, which is why we write the test down before we run it and keep the Elo rule frozen.

None of the shot, goalie or special-teams adjustments lifted the hit rate of our 62% Elo tier either. For the record of that tier this season, see NHL moneyline and puck line. Nothing on this page is a reason to back or fade any team.

Frequently asked questions

Does the penalty kill predict who wins NHL games?

Barely. Over five seasons, adding power-play and penalty-kill percentages to our Elo improved predictions in only one season. Penalty-kill percentage swings a lot by chance: even across a full 82-game season, a team's odd and even games agree with a correlation of only about 0.30.

Does the starting goalie matter in the NHL?

A little, but we could not prove how much. A team starting its backup won about 1.5 points less often on the fairest measure, and that range includes zero. Adding goalie information to a prediction model made it less accurate, not more, and goalie save quality was weakly repeatable from one stretch to the next.

What is the best predictor of NHL results after wins and losses?

Shot quality is the strongest lead we found. Adding the 5v5 expected-goals share to our Elo improved its predictions in four of five seasons. But in our single forward test season the gain was small and not statistically clear, so we call it a lead, not a finding.

The week ahead, once a week

Last week's graded record — wins and losses — and the plays for the week ahead. One email, no daily noise, unsubscribe in a click.

Educational content only. Research results are historical, from our own data, and describe what happened, not what will happen. Must be 21+ and in a jurisdiction where sports betting is legal. If gambling is a problem for you or someone you know, call 1-800-GAMBLER or visit ncpgambling.org.