Sports

NHL betting models built around fair price

Hockey is decided by a handful of goals. That single fact drives the market structure, the variance a model has to live with, and the reason the three-way regulation market exists at all.

Updated October 23, 2026 · 12 min read

Analytics only. quant+ never places bets or holds funds.

Low scoring turns every goal into a large event

In a sport where the average team scores a small number of times, the difference between a good performance and a bad one can be one bounce. That is not a flaw in hockey analysis. It is the central fact a hockey model has to be built around.

Skew and the empty net

Trailing teams pull the goalie, which converts a one-goal deficit into either an equaliser or an additional goal against. The result is that a meaningful share of goals are scored into an undefended net, and the final score is a less faithful record of the play than the shot count. A model that treats every goal as equal information is treating a tactical decision as a hockey event.

One goal is never safe

A one-goal lead in the third period is not a settled game in the way that a two-touchdown lead is in football. The trailing team can equalise on a single shot with the goalie pulled, and the resulting overtime is a short, high-variance period that frequently ends in a shootout. Both of those stages are close to coin flips.

Ties are structural, not rare

Hockey is the only major North American team sport that routinely reaches the end of its regulation period level. A regulatory draw that then resolves in a three-on-three period or a shootout is a genuinely different competitive state, which is why the market splits it out into its own price rather than burying it.

The three-way market is a sixty-minute opinion

The regulation market asks who wins the first sixty minutes and offers a draw as a third option. The two-way market asks who wins the game, including whatever happens after the horn. They are not competing quotes on one question. They are quotes on two different questions that overlap for most of the game.

The practical value of the split is that it separates skill from resolution. A model that is confident about which team is better over sixty minutes can express that in the regulation market without being forced to include a five-minute three-on-three period and a shootout competition, both of which compress the difference between teams sharply. If the model is right about the sixty minutes, the regulation price is where that shows.

The two markets also discipline each other. If a model produces a regulation price and a two-way price whose implied probabilities are not consistent with a sensible value for the overtime segment, one of the two outputs is wrong. That internal check costs nothing and catches a class of error that is otherwise easy to miss, because both markets look individually plausible on their own.

  • Three-way regulation. A market with three outcomes: home win in regulation, away win in regulation, and the tie that sends the game to overtime. Because it excludes the extra period, the two win prices are longer than on the two-way market, and the tie is a genuine third outcome rather than a void. It is the market that most directly expresses a model's honest opinion about the sixty minutes, before the coin-flip portion of overtime is introduced.
  • Two-way moneyline. Home or away to win the game by any means, including overtime and a shootout. The extra period resolves what the three-way market leaves open, so the two markets are not two opinions about the same question. They are different questions, and the difference between them is the price of the overtime and shootout segment.
  • Puck line and totals. The puck line is a fixed handicap, conventionally one and a half goals, with the favourite typically priced at plus money. Totals hover in a narrow band around five and a half to six and a half goals, which means a single goal is a large fraction of the number and pushes and half-goal differences carry real weight.
  • Goalie and skater props. Save totals, goals against, and skater shot and point props. These sit on top of the same variance structure and inherit it: a goalie prop depends on how many shots he faces, which depends on the game state in ways that are not fully forecastable.

The goalie on the night is not always the goalie on the chart

Baseball publishes its starting pitchers days ahead. Hockey does not. The starting netminder is often confirmed only shortly before the game, and the uncertainty around it is a first-class modelling problem rather than a data gap.

Three states of knowledge

  • Unconfirmed: the price is really a probability-weighted blend over possible starters, and treating it as a single-goalie price overstates confidence.
  • Confirmed: the model can price the actual matchup, and the line has usually already begun to move.
  • Changed late: a warmup injury or a coach decision replaces the starter, and a market that was not action-regardless has to be re-read rather than assumed.

The backup goalie question is where this becomes interesting. Teams manage starter workload deliberately, and the decision to start the backup is frequently tied to the schedule rather than to form: the second night of a back-to-back, the tail of a road trip, or a run of three games in four nights. That means the backup substitution is partly predictable from the calendar even when it is not announced.

A model should therefore carry the starter as an uncertain input with a distribution, not as a known quantity, and it should update that distribution when the schedule context makes a backup start likely. The cost of getting this wrong is asymmetric: assuming the number one goalie and getting the backup, on a game where the price has not yet adjusted, is the kind of error that looks like a large edge and is really a stale model.

There is a corollary for reading prices. A line that appears to be several percentage points away from a model's number is often not a disagreement about hockey at all. It is frequently a disagreement about who is starting, and the model is the side that does not know yet.

Six inputs, with the schedule doing more work than in any other sport

Hockey's inputs are less about individual talent than about context: who is on the ice, who is in net, and what the team did last night.

The confirmed starting goalie

Who starts in net is the largest single swing factor in a hockey price, and it is not published the way a baseball rotation is. Coaches name a starter close to the game, sometimes on the day, and a backup starting instead of the number one can move a line materially. Until the confirmation is out, the market is pricing a probability over starters rather than a single one.

Back-to-backs and the schedule

Teams play on consecutive nights and travel between them. A team on the second night of a back-to-back, particularly after travel, is a different proposition from the same team rested, and this is one of the few schedule effects in any sport that is both large and known in advance.

Line combinations and defensive pairs

Hockey is played in shifting groups rather than by a fixed lineup, so who skates with whom determines how much of the game a given player is on the ice for. Injuries and coach decisions reshuffle the combinations, and the reshuffle changes matchup structure even when the roster is largely intact.

Shot volume and quality

Shots are far more numerous than goals, so they carry more signal per game. Modelling both the number of shots and their quality is a way of stabilising a statistic that is otherwise too rare to be informative in a single game. A team that is being outshot is telling you something that the scoreline may not yet reflect.

Rest, travel and time zones

Fatigue effects in hockey are well documented and schedule-driven. A west-coast trip, a long road stretch, or a game following a late finish all appear in the schedule before they appear in the price, which makes the schedule one of the more tractable inputs available.

Special teams

Power play and penalty kill are discrete, high-leverage game states that behave differently from even-strength play. Because penalties are partly discretionary, the number of opportunities varies, which means a special teams advantage is a rate applied to an uncertain number of opportunities rather than a fixed addition.

A narrow band where half a goal is a lot

Hockey totals cluster tightly, which makes the market sensitive to differences that would be noise on a higher-scoring sport. The table below sets out which input is expected to act on which side of the number.

Input Mechanism Expected direction
Backup goalie start A less experienced netminder in place of the regular starter, on a schedule that already suggests one is coming. Raises the projected total
Back-to-back, second night Fatigue and, frequently, a backup start in the same game, which compounds rather than offsetting. Raises the projected total
Long road trip, final game Accumulated travel and time-zone change reduce intensity at both ends, which tends to suppress rather than inflate. Lowers the projected total
Two strong starters confirmed The most likely scenario for a low number, since both teams are at full strength in net and neither side is likely to be chasing. Lowers the projected total
Evenly matched, low-event teams A game that stays close is more likely to reach the pulled-goalie state, which adds empty-net goals that are recorded like any other. Weak upward pressure on the final score

Directions are structural expectations about sign, not measured effect sizes. No numeric adjustment is claimed for any row and none should be read into the table.

Three ways a hockey model looks better than it is

The same traps that catch any sports model are more dangerous in hockey, because the low-scoring structure lets a small number of results look like a pattern.

Modelling the wrong goalie

A backtest that uses the goalie who actually played has seen the future. The reproducible version has to use the information available before the game, which frequently means a mix of possible starters. The gap between those two backtests is usually the difference between a profitable model and a losing one.

Ignoring the pulled goalie

Empty-net goals are a consequence of game state, which means they are correlated with the score rather than independent of it. A model trained on final scores without accounting for that will overrate late scoring and misprice totals in close games, precisely the games it is asked about most.

Treating overtime as regulation

A two-way moneyline and a regulation market are not interchangeable, and a backtest that mixes them, or that settles a three-way bet on the final score, will report results that do not correspond to any market that exists. The settlement rule has to be applied the way the book applies it.

The general method does not change with the sport: evaluate in point-in-time order, score against closing prices, and insist on a large sample. See the model building guide.

Where an NHL model fails

Hockey may be the sport where a correct probability looks most wrong most often. A small number of goals per game means the outcome distribution is coarse, and a measured edge is very hard to distinguish from noise over anything short of a long run of games. Anyone presenting a hockey model with a short, strong record is presenting a small sample.

The market is also efficient in the places that are easiest to model. The two-way moneyline on a well-rested team with a confirmed starter is priced by people who know the sport well, and a large apparent edge there should be treated as a suspicion about the model rather than a discovery. The areas where a careful process has more room are the ones that require a probability over uncertain inputs: unconfirmed goalies, schedule-driven fatigue, and totals in games where the pulled-goalie state is likely.

Finally, a model cannot see the room. A coach's decision to rest a player, a mid-game injury, or a fight that shifts momentum are not forecastable from a spreadsheet. The honest position is that the model prices the part of the game that is measurable and that the rest is variance.

quant+ publishes the research surface. It does not place bets, does not hold funds, and has no account connection to any sportsbook.

Scan the hockey board · Compare prices across books · Open the terminal

Price the sixty minutes, then the rest

Separate what the model knows about regulation from what overtime resolves, carry the goalie as an uncertain input, and check the schedule before the price.

Analyze the NHL slate · See the model surface

Questions people ask

What is an NHL betting model?

A system that converts team strength, the confirmed or probable starting goalie, line combinations and the schedule into a distribution over goals for each side, and from that into a fair moneyline, regulation three-way price, puck line and total. Because hockey scores so few goals, the distribution matters as much as the mean, and the model's output is only useful once it has been expressed as a probability and compared with a margin-removed price.

What is the three-way regulation market in hockey?

A market with three outcomes: the home team wins in regulation, the away team wins in regulation, or the game is level after sixty minutes and goes to overtime. It excludes the extra period, so it prices only the regulation result. It differs from the two-way moneyline, which settles on the final result including overtime and a shootout, and the difference between the two prices reflects the value of that extra segment.

Why does the starting goalie matter so much?

Because the goalie is on the ice for the entire game and the position has more influence on the scoreline than any skater. Unlike a baseball rotation, NHL starters are often confirmed only shortly before the game, and a backup starting in place of the regular netminder can move a line noticeably. Until the confirmation is out, any model price is really a blend over possible starters.

How do back-to-backs affect NHL odds?

Teams sometimes play on consecutive nights, often with travel between the two games, and the second night is measurably harder. The effect compounds when the schedule also points to a backup goalie starting. Crucially, the schedule is known well in advance, so this is one of the few context inputs that can be applied before the market has fully absorbed it.

Why are NHL totals so low and so tight?

Because hockey games produce a small number of goals per team, totals cluster in a narrow band and half a goal is a large fraction of the number. That makes the total unusually sensitive to inputs that would be rounding errors in a higher-scoring sport, and it makes the push probability on a whole-number line a meaningful part of the pricing rather than an edge case.

Can an NHL model beat the closing line?

Occasionally, on the markets that require a probability over uncertain inputs, such as games with an unconfirmed goalie or on the second night of a back-to-back. On the most liquid two-way moneylines, a large apparent edge is more likely to indicate a modelling error or a stale goalie assumption than a genuine mispricing. As in every sport, the test is a point-in-time evaluation against closing prices over a large sample.

Sources

  1. What a sports betting model outputs | Rithmm, rithmm.com
  2. Three-way betting (attempted) | Wikipedia, en.wikipedia.org
  3. Puck line (attempted) | Wikipedia, en.wikipedia.org