Does the Starting Goalie Change Who Wins an NHL Game? We Tested 6,990 Games

By: Steve Pugh, Founder, Versus Sports Simulator
November 4, 2026

When a team's usual goalie sits and the backup starts, does that change which team is likely to win the game? Versus rates each NHL team from its game results, and those results already include the team's goalies, starters and backups together. But our predictions do not adjust for which goalie starts on a given night, so we tested whether they should.

We rated every goalie who played in the NHL from 2015-16 through 2025-26. Then we went back through five full seasons, 6,990 games, and compared our regular predictions with the same predictions after adding each game's actual starting goalie. For each set, we counted the games where the prediction picked the team that won.

Our regular predictions picked the winner of 4,087 games. With the starting goalie added, they picked the winner of 4,062, which is 25 fewer. The starting goalie does matter, but only a little. A backup costs his team about a tenth of a goal per game, and a tenth of a goal rarely changes which team should be favored.

Does the Versus NHL Model Account for the Goalie?

Versus rates every NHL team from its game results, with one number for offense and one for defense, and predicts each game from those two numbers. The defensive number is built from the goals a team has actually allowed. So a team's defensive rating already includes its goalies, starters and backups together, in proportion to how often each one has played.

But our predictions do not adjust for which goalie starts on a given night. If a team's usual goalie has started 60% of its games and the backup starts tonight, the team's defensive rating still describes the usual mix of the two. And backups start often. From 2021-22 through 2025-26, someone other than a team's usual goalie started 4 of every 10 of its games.

How We Tested the Starting Goalie

We took every goalie appearance from the NHL's own game logs, regular season and playoffs, from 2015-16 through 2025-26. That comes to 31,011 appearances by 243 goalies across 14,508 games.

Each goalie's rating counts how many more saves he made than an average goalie would have made on the same shots, which hockey analysts call goals saved above average. And recent games count more than old ones. A goalie with few games played starts with a rating a little below average, and his own record takes over as he faces more shots. We start new goalies slightly below average because that fit past results better than starting them at average.

For every game, we compared the starting goalie with the mix of goalies already included in his team's defensive rating. If the starting goalie was better than that mix, we lowered the other team's predicted goals. If he was worse, we raised the other team's predicted goals. Nothing else in the prediction changed.

The goalie rating has a few settings, such as how fast old games fade. We chose those settings using only 2016-17 through 2020-21, then locked them before testing on 2021-22 through 2025-26. Our regular predictions in this test are today's Versus model rerun on the 2021-22 through 2025-26 seasons, predicting each game only from the games played before it.

And the test used the goalie who actually started each game. Our predictions are made overnight, before most starting goalies are confirmed, so the test used better goalie information than any overnight prediction could have. That gave the starting goalie its best possible chance to improve our predictions.

Does Adding the Starting Goalie Pick More Winners?

Adding the starting goalie does not pick more winners. Over the five test seasons, our regular predictions picked the winner of 4,087 of 6,990 games. With the starting goalie added, they picked the winner of 4,062. And the version with the starting goalie picked fewer winners in all five seasons.

Adding the starting goalie changed the predicted winner in only 473 games, about 1 in 15. In those 473 games, the version with the starting goalie was right 224 times and our regular predictions were right 249 times. A fair coin flipped 473 times splits at least that unevenly about 27% of the time, so honestly I think the drop of 25 winners is mostly luck. Either way, adding the goalie did not pick more winners.

We also tested two simpler versions, with their settings locked before the test as well. One applied the same penalty to a team every time its usual goalie sat, and it picked 17 fewer winners than our regular predictions. The other used the goalie rating only in games a backup started, and it picked 2 fewer. So none of the three versions picked more winners than our regular predictions.

How Much Is a Backup Goalie Worth in the NHL?

A backup goalie costs his team about a tenth of a goal per game. From 2021-22 through 2025-26, teams allowed about 0.1 more goals per game when someone other than their usual goalie started, after accounting for the opponent. In the five seasons before that, the difference was about half as big. We count a team's usual goalie as the one with the most starts for that team so far that season.

The goalie ratings measure more than usual goalie versus backup, since some backups are better than some usual goalies. But even using each goalie's own rating, a team's starting goalie typically moved the other team's predicted goals by only 0.05, and 9 times out of 10 he moved them by less than an eighth of a goal. In the typical game, adding the goalie moved a team's chance of winning by 1 percentage point, and in 90% of games it moved that chance by 3 percentage points or less.

So the starting goalie has a real effect, but a small one. When a backup starts, his team really does allow slightly more goals.

Why the Starting Goalie Rarely Changes the Predicted Winner

The predicted winner only changes when the goalie's effect is bigger than the predicted gap between the two teams. Say we have Boston as a 55% favorite. If Boston starts its backup, that might move Boston to 53%. Boston is still the predicted winner, only by a smaller margin.

The 473 games where the predicted winner did change were games our predictions already had as very close. The typical favorite in those games was at 51%, and none was above 58%. Switching from a 51% favorite to the other team at 51% is almost exactly a coin flip, so over 473 games luck decides most of which version comes out ahead.

Even a perfect goalie rating would not have shown a clear gain. If every one of our goalie ratings had been exactly right, the version with the starting goalie should have picked the winner in about 243 of those 473 games, compared with 236 or 237 for a coin flip. But by pure chance, 473 coin flips land heads anywhere from about 215 to 258 times, 95% of the time. So a real edge that small can't be told apart from luck in five seasons.

So the next time you see a backup starting in a game you care about, check how far above 50% our favorite is first. In five seasons, adding the goalie never changed the predicted winner of a game where our favorite was above 58%.

Why NHL Games Are So Hard to Predict

NHL games are close. From 2021-22 through 2025-26, 40% of NHL games ended with a one-goal margin and 21% went past regulation into overtime or a shootout. Home teams won 53.6% of the time. Teams averaged 3.1 goals a game, so one bounce or one deflection decides a lot of games.

And those close games show up in our predictions. In two games out of three, our favorite had less than a 60% chance to win, and the typical favorite had a 56% chance. So a favorite at 56% should lose 44% of the time. But our win percentages match what actually happened.

Over the five seasons our regular predictions picked the winner of 58.5% of NHL games, from 62.5% in 2021-22 down to 54.6% in 2025-26, the hardest of the five seasons to predict. And I'd rather publish a 56% favorite that wins 56% of the time than a 70% favorite that wins 60% of the time. Every NHL game's win percentage for the coming week is on our NHL predictions page.

The Math Behind the Test

The Goalie Rating

For goalie \(g\), every past game \(i\) contributes his saves above what a league-average goalie would have made on the same shots:

\[ r_g \;=\; \frac{\sum_i w_i\,\bigl(\text{saves}_i - \bar{s}_i\,\text{shots}_i\bigr) \;+\; K\,\mu_0}{\sum_i w_i\,\text{shots}_i \;+\; K} \]

Here \(\bar{s}_i\) is the league's save percentage over the 365 days before game \(i\), and \(w_i = 2^{-a_i/365}\) halves a game's weight for every year since it was played, with \(a_i\) its age in days. \(K\) is a prior of 1,500 shots at \(\mu_0\), a save percentage 0.004 below league average. A goalie with little history starts slightly below average, and his own record takes over as his shots add up. Only games before the day being predicted count.

Comparing the Starter With the Team's Usual Goalies

A team's defensive rating is fitted on its games so far this season plus a preseason starting point carried over from last season. So the starter is compared with the goalie mix inside that rating:

\[ m_T \;=\; \sum_g \pi_{T,g}\; r_g \]

where \(\pi_{T,g}\) is goalie \(g\)'s share of team \(T\)'s ice time this season, with last season's mix counted as ten extra games. The adjustment, in goals, is

\[ \Delta_T \;=\; \bigl(r_{\text{starter}} - m_T\bigr)\cdot \widehat{\mathrm{SA}}_T \]

with \(\widehat{\mathrm{SA}}_T\) the shots the team has allowed per game this season, pulled toward the league average by ten games. The opponent's predicted goals drop by \(k\,\Delta_T\). The team's own predicted goals don't move.

From Goals to a Win Percentage

Versus turns the predicted goal difference \(d\) between two teams into a win percentage with a normal curve 2 goals wide:

\[ p \;=\; 0.01 + 0.98\,\Phi\!\left(\frac{d}{2}\right) \]

At an even game the slope of that curve is \(0.98\,\varphi(0)/2 \approx 0.196\) per goal, so every tenth of a goal is worth about 2 points of win percentage. The predicted winner is whichever team has \(p\) above 50%, so an adjustment can only change the predicted winner of a game whose \(|d|\) is smaller than the adjustment itself.

Choosing the Settings Without Seeing the Test

Four settings were searched: the prior \(K\) (750, 1,500, 3,000 or 6,000 shots), the half-life (one year, two years, or no fading at all), the prior \(\mu_0\) (league average, or 0.004 below it) and a scale \(k\) on the adjustment (0.5, 1 or 1.5). That makes 72 combinations. The one kept had the lowest log loss on 2016-17 through 2020-21,

\[ \text{log loss} \;=\; -\frac{1}{n}\sum_{j=1}^{n} \ln p_j\bigl(\text{actual winner}\bigr) \]

which rewards win percentages that match what happened. The combination with the lowest log loss was \(K = 1{,}500\), a one-year half-life, \(\mu_0\) at 0.004 below average, and \(k = 1\). Then that combination was run once on 2021-22 through 2025-26. Looking back afterward, all 72 combinations landed between 4,011 and 4,092 winners on the test seasons, against 4,087 for our regular predictions. So even the best combination, picked after seeing the results, would have added only 5 winners.

The Coin-Flip Check

The predicted winner changed in \(n = 473\) games. If the goalie added nothing, the number of those games the version with the starting goalie gets right would behave like a fair coin:

\[ X \sim \mathrm{Binomial}\!\left(473,\ \tfrac{1}{2}\right) \]

A count like that lands in this range 95% of the time:

\[ 236.5 \pm 1.96\sqrt{473/4} \;\approx\; 215 \text{ to } 258 \]

The version with the starting goalie got \(X = 224\). A split at least that far from even happens with probability

\[ \Pr\bigl[\,|X - 236.5| \ge 12.5\,\bigr] \;\approx\; 0.27 \]

The win percentages \(q_j\) from the version with the starting goalie say what a perfect rating should have produced on those games, an expected net gain of \(\sum_j (2q_j - 1) \approx 13.8\) winners, or about 243 of the 473. That figure sits well inside the coin's range too, which is why five seasons can't separate a real edge this small from luck. The win percentages from that version also came slightly closer to the results than our regular predictions did, by about one standard error, which is also the size luck produces.

The Goalie Effect Is Real

Winners aside, the goalie adjustment does predict goals. Regress each team's goals allowed, minus what our regular predictions expected, on its adjustment:

\[ \text{allowed}_T - \widehat{\text{allowed}}_T \;=\; \alpha + \beta\,\Delta_T + \varepsilon \]

A perfect adjustment gives a slope of minus one. Across the 13,822 team games of the test seasons, leaving out each team's first game of a season, the slope came out at minus 0.78, with a standard error of 0.18. So the goalie ratings carry real information about goals. But that effect is so small, a median of 0.05 goals per team per game, that it rarely changes who wins.

Frequently Asked Questions

Does the starting goalie matter in NHL predictions?

The starting goalie matters a little. Over 6,990 NHL games from 2021-22 through 2025-26, adding the actual starting goalie to Versus Sports Simulator's predictions changed the predicted winner in 473 games and produced 25 fewer correct predictions overall. A backup goalie costs his team about a tenth of a goal per game, which moves its chance of winning about 2 percentage points, too little to change the favorite in most games.

How much is a backup goalie worth in the NHL?

A backup goalie costs his team about a tenth of a goal per game. From 2021-22 through 2025-26, NHL teams allowed about 0.1 more goals per game when someone other than their usual goalie started, after accounting for the opponent. In the five seasons before that, the difference was about half as large.

How accurate are NHL game predictions?

Rerun on 2021-22 through 2025-26, the current Versus Sports Simulator model picked the winner in 58.5% of NHL games (4,087 of 6,990), ranging from 62.5% in 2021-22 to 54.6% in 2025-26. Its win percentages match the results. Favorites given 60% to 64% won 60.5% of the time, and favorites given 70% or more won 73.2%.

What percentage of NHL games are decided by one goal?

About 40% of NHL games are decided by one goal. From 2021-22 through 2025-26, 2,790 of 6,990 NHL games ended with a one-goal margin, and 21% went past regulation into overtime or a shootout.

What is goals saved above average?

Goals saved above average measures how many more saves a goalie made than a league-average goalie would have made on the same shots. A goalie who faces 1,000 shots with a save percentage 0.010 above the league average has saved about 10 goals above average.

Sources

1. NHL Stats: Goalies. National Hockey League.