Learn
The Puck Luck Test: an NHL betting sample size table for every edge
NHL betting sample size is where most hockey models get judged too early. The Puck Luck Test shows how much of a hockey record is luck, how often a hypothetical model with a real edge still finishes a full 1,312-game season below break-even, how many bets each edge needs, and why beating the closing line is the better test in the meantime.
One NHL season is 1,312 regular-season games (32 teams, 82 games each), and for judging a hockey betting model that still isn't much. The Puck Luck Test is the three moves from the video, in the same order, and the table below is the part you will want to keep.
- Know how much of hockey is luck. One study put luck at about 53% of an NHL team's season record, against about 12% in the NBA (The Conversation).
- Run the season math. Work out how often a model with a real edge still finishes a full season below break-even, and how many bets the edge needs before luck stops being the likely explanation.
- Check the close. Until you have that many bets, judge the model by whether its prices beat the closing line, not by its record.
Step 1. Know how much of hockey is luck
In The Conversation, Mark Robert Rank used true score theory to split a team's record into skill and luck, and found luck accounts for about 53% of an NHL team's season record, compared with about 12% for an NBA team (The Conversation). Low scoring is a big part of it: one bounce decides far more hockey games than basketball games.
That matters for a betting model in two ways. A model that is right about the teams can still lose a lot of single games, and a model that is wrong can win a lot of them. So the record of a hockey model carries more noise per bet than the same record in basketball, and the sample has to be bigger before it says anything.
Step 2. Run the season math
Start with the bar. At -110 you risk 110 to win 100, so the break-even win rate is 110 divided by 210, or 52.38% (moneyline odds). Hockey sides are usually priced as moneylines with a different price on each team, and the bar moves with the price, so the table uses -110 to keep the arithmetic simple. Puck lines and totals often sit near that price.
Then count the season. The league has 32 teams playing 82 games each, which is 1,312 regular-season games (2025-26 NHL season; 32 times 82 divided by 2 is our arithmetic). A model that bet every one of them would have the biggest sample a single season can give.
The table is hypothetical. For a model whose true win rate at -110 is the number on the left, it shows how often the record would still sit below break-even after a given number of bets, and how many bets it takes before the record sits about two standard errors above the bar. The math is the normal approximation to the binomial distribution.
| True win rate at -110 (hypothetical) | Below break-even after 100 bets | After 300 | After 1,312 (one season, every game) | Bets for a 2-standard-error margin | Seasons of every game |
|---|---|---|---|---|---|
| 53% | 45% | 42% | 33% | about 26,000 | about 20 |
| 54% | 37% | 29% | 12% | about 3,800 | about 2.9 |
| 55% | 30% | 18% | 3% | about 1,440 | about 1.1 |
| 56% | 23% | 10% | under 1% | about 750 | about 0.6 |
| 57% | 18% | 5% | under 1% | about 460 | about 0.4 |
| 58% | 13% | 2% | under 1% | about 310 | about 0.2 |
The bets-needed column uses the standard error of a win rate, the square root of p times one minus p over n, and asks for the gap to break-even to be twice that size (binomial distribution).
Read the 53% row slowly, because that is what a real but small edge looks like. A model that good, betting every game of a season, still finishes the year below break-even about one time in three, and proving the edge from the record alone takes about 20 seasons. A 55% model drops the chance of a losing season to about 3 in 100 (binomial distribution), which is why bigger edges show up faster. Edges that big are rare and rarely last, so most of the time you are in the top rows of the table.
The same math cuts the other way. In the same hypothetical approximation, a model with no edge at all, a true 50%, finishes a 100-bet stretch at 55% or better about one time in six (binomial distribution). A good month is not evidence by itself.
Step 3. Check the close while you wait
You can't wait 26,000 bets to find out whether a model is any good. The closing line is the market's last price before the game, and markets are hard to beat: Wunderlich and Memmert note that betting odds are frequently found to outperform mathematical models. If the prices a model takes keep beating the close, that is a signal you can measure on every single bet, long before the record settles.
Beating the close also doesn't require being a better forecaster than the market on every game. Hubáček and Šír show a model can profit while predicting worse than the market, as long as its errors are not correlated with the market's. What matters is where its prices disagree with the market and whether the market later moves toward them.
The routine:
- Log every bet with the price you took and the closing price for the same side.
- Mark each one beat, matched or missed. Our closing line value calculator does the math.
- After every 50 bets, look at the share you beat and the average size of the beat, not at the record.
A season log you can copy
| Month | Bets | Record | Win rate | Bets that beat the close | Share beating the close |
|---|---|---|---|---|---|
| October | |||||
| November | |||||
| Season total |
Write one rule at the top before the first game: which games the model bets, at what price, and how much you stake. Don't change it mid-season.
What to watch out for
- Judging by the record. In the hypothetical table, a true 53% model finishes 100 bets below break-even about 45% of the time (binomial distribution). A losing month and a winning month can come from the same model.
- Changing the rule after a losing run. That resets your sample to zero.
- Testing many models and keeping the best season. With this much luck in hockey, one of several no-edge models will often look great for a year.
- Treating goals like basketball points. Low-scoring games swing on one bounce, which is the reason the luck share is so high in the first place.
- Stake creep. A hot start talks you into bigger bets on a skill you haven't proven. Keep the unit fixed for the whole test.
How to tell it worked
By the end of a season you won't have proof from the record, and that's expected. What you should have is a clear closing-line number across a few hundred bets and a record that sits inside the range the table predicts for the edge you think you have. If the prices rarely beat the close, the record was probably luck in either direction. Every Quant+ model publishes a scorecard with how often it beats the close, walk-forward by season, so you can run the same test on ours.
21+. For informational purposes only. Nothing is guaranteed. If gambling stops being fun, call 1-800-GAMBLER.
Questions people ask
How much of an NHL team's record is luck?
One study using true score theory put it at about 53% of an NHL team's season record, against about 12% in the NBA (The Conversation).
Is one winning NHL season proof that a betting model works?
No. In a hypothetical where a model truly wins 53% at -110 and bets all 1,312 regular-season games, it still finishes below break-even about one time in three.
How many bets does a hockey model need?
For the record to sit two standard errors above the 52.38% break-even, a true 55% model needs about 1,440 bets and a true 53% model about 26,000. These are hypothetical numbers from the normal approximation.
What can I check before I have that many bets?
Closing line value. If the prices a model takes keep beating the closing line, that is evidence you can measure on every bet, long before the win-loss record settles.
Sources
- Why luck plays such a big role in hockey, The Conversation (Mark Robert Rank)
- 2025-26 NHL season, Wikipedia
- Moneyline odds, Wikipedia
- Binomial distribution, Wikipedia
- The Betting Odds Rating System, PLOS ONE (Wunderlich, Memmert)
- Beating the market with a bad predictive model, arXiv (Hubáček, Šír)
