NHL Betting Analytics: A Machine Learning Approach to Hockey

Unlike most North American sports, NHL moneylines come in two distinct formats. The 3-way moneyline covers regulation time only — you can bet Team A, Team B, or the draw. If the game goes to overtime or a shootout, the draw wins. The 2-way moneyline (also called “including overtime”) is the more common retail bet: pick which team wins the game outright, regardless of whether it takes OT or a shootout to get there.

The 3-way market is where sharper bettors often find value. Roughly 23-25% of NHL games go to overtime, meaning the draw is a legitimate outcome with meaningful probability. Books price the 3-way draw anywhere from +260 to +340 depending on how evenly matched the teams are. Because recreational bettors overwhelmingly prefer the 2-way market (they want to pick a winner, not bet on a tie), the 3-way line receives less action and less sharpening pressure.

NHL moneylines are significantly tighter than MLB moneylines. While baseball regularly produces -250 or -300 favorites, hockey games rarely see moneylines beyond -200. The reason is structural: parity in the NHL is higher than in any other major North American league. In any given season, the gap between the best and worst teams is compressed because a hot goaltender can carry an otherwise mediocre roster deep into a playoff run. This compression means that value on heavy favorites is rare, and underdogs cover more often than in other sports.

The puck line is hockey's version of the point spread, fixed at -1.5 goals for the favorite and +1.5 for the underdog. In a sport where the average game produces 6.0-6.4 total goals, a 1.5-goal margin is enormous. About 25% of NHL games are decided by exactly one goal in regulation, meaning the puck line and moneyline disagree roughly a quarter of the time.

Because the puck line is fixed (unlike NFL or NBA spreads that move), the odds around it adjust instead. A moderate favorite might be -1.5 at +150, while a heavy favorite might be -1.5 at -120. The alternate puck line market extends this to -2.5 and +2.5, but liquidity drops sharply beyond the standard 1.5.

One key strategic consideration: empty-net goals. In the final two minutes of a game, the trailing team almost always pulls their goaltender for an extra skater. This creates a high probability of either a tying goal or an empty-net goal that extends the margin. Models that account for empty-net goal probability in their margin-of-victory estimates gain an edge on puck line pricing.

NHL game totals typically sit between 5.5 and 6.5, reflecting the combined expected goals for both teams. The goaltender matchup is the dominant factor in totals pricing — a game between Connor Hellebuyck and Igor Shesterkin might open at 5.5, while a game featuring two backup goalies could open at 6.5 or higher.

Unlike MLB where park factors create huge variation in run environments (Coors Field vs. Oracle Park), NHL arena effects on scoring are relatively small. The bigger factors are team pace (shots per game), power play opportunities, and goaltender quality. Special teams efficiency matters enormously: a team with a 25% power play converts at nearly double the rate of a team at 14%, and penalty frequency varies meaningfully by matchup and referee crew.

NHL player props are the fastest-growing segment of hockey betting and the area with the most exploitable inefficiency. The primary prop markets include:

The NHL prop market is thinner than MLB or NBA props — fewer sportsbooks offer them, limits are lower, and the lines are less sharp. This is precisely what creates opportunity for models. When a market is inefficient, the gap between true probability and book-implied probability widens.

NHL live betting updates continuously through each of the three 20-minute periods and overtime. Live markets include adjusted moneylines, running game totals, period-specific totals, and next goal scorer. The pace of NHL scoring events — goals, power plays, penalties — creates frequent line movements that models processing real-time data can exploit.

A particularly interesting live dynamic is the power play. When a team goes on a 5-on-4 advantage, the expected goals spike temporarily. If the live total has not adjusted for an upcoming power play (or a double minor penalty), there is a brief window where the over becomes +EV before the book recalibrates. Similarly, if a starting goaltender gets pulled mid-game for poor performance, the live moneyline and totals should shift dramatically — but books are sometimes slow to react.

Converting NHL odds to implied probability works identically to other sports:

NHL moneyline vig is generally tighter than MLB (3-4% vs. 4-5%) because the odds spread is narrower. However, NHL player prop vig is often wider — 6-10% on standard markets, sometimes 10-14% on less liquid props like hits or blocked shots. The model needs to clear a higher hurdle on props, but the inefficiency in pricing is also larger.

In no other sport does a single roster announcement move the line as much as goaltender confirmation in hockey. Starting goaltenders are typically confirmed by teams or reporters 10-12 hours before puck drop, but sometimes not until morning skate (5-6 hours before game time). When an unexpected backup starts, moneylines can shift 20-40 cents in minutes.

Consider the difference: if the New York Rangers start Igor Shesterkin (.925 save percentage, Vezina Trophy winner) versus their backup (.905 save percentage), the expected goals against changes by 0.6-0.8 goals per game. On a 30-shot night, the gap between a .925 and .905 save percentage is roughly 0.6 additional goals allowed. That is an enormous swing in a sport where the average winning margin is 1.5 goals.

Models that incorporate real-time goaltender confirmation — and can recalculate predictions instantly when goalie news breaks — gain a significant timing edge over bettors who rely on pregame analysis that assumed the starter would play.

NHL home ice advantage is real but modest — home teams win approximately 54% of games, one of the smallest home advantages in major professional sports. The advantage comes from last change (the home team gets to make the final line matchup), favorable faceoff placement, and crowd energy. Some arenas have slightly larger or smaller ice surfaces at the international level, but all NHL rinks are standardized at 200 by 85 feet.

Books bake home ice into their lines, typically adjusting by 5-8 cents on the moneyline. The question for models is whether the adjustment is accurate. Historical data suggests that home advantage fluctuates by team — some teams have consistently stronger home records (often correlated with altitude, like Colorado) while others show minimal split. A model that uses team-specific home/away adjustments rather than a league-wide average captures this nuance.

The NHL schedule is grueling — 82 games in roughly 180 days, with frequent back-to-back games (a team plays one night and again the next night, often in a different city). Back-to-back scheduling has a measurable impact on performance: teams playing the second game of a back-to-back win at roughly a 2-3% lower rate than normal, and their goals against average increases by approximately 0.2-0.3 goals.

More importantly, teams almost always start their backup goaltender on one leg of a back-to-back. This compounds the fatigue effect with the goaltender quality effect described above. Books generally account for this, but the speed and accuracy of their adjustment varies. When a team is on the second night of a back-to-back, traveling across time zones, with their backup in net — and the book has only partially adjusted — the model finds value.

In a high-scoring sport like basketball (110+ points per team), the better team wins most of the time because a large number of scoring events allow skill to dominate luck. In hockey, where a game might produce only 5-6 combined goals, a single lucky bounce, deflection, or bad-angle goal can flip the outcome. The result is more upsets, tighter moneylines, and lower maximum model confidence.

To put this in context: in the NBA, a model might identify a game where one team has a 75-80% win probability. In the NHL, the practical ceiling for model confidence is closer to 62-65% for any single game. Even the best team in the league loses 35% of its games in a given season. This means that NHL model outputs should be interpreted differently — a 58% confidence level in hockey is the equivalent of a much higher confidence level in basketball, relative to the achievable range.

Goal scoring in hockey closely follows a Poisson distribution — a probability distribution that models the number of independent events occurring in a fixed time interval. The key Poisson assumptions map well to hockey: goals are relatively rare, occur independently of each other (mostly), and have a roughly constant rate within a game.

The Poisson framework is especially useful for game totals. If Team A has an expected goals rate of 3.1 and Team B has 2.8, you can model the combined total as a Poisson variable with lambda = 5.9 and calculate precise probabilities for Over/Under 5.5, 6.5, or any other line. This is mathematically cleaner than trying to directly classify “over” or “under” as a binary outcome.

Perhaps the most frustrating challenge in NHL modeling is the “hot goalie” phenomenon. A goaltender can, in a single game, suppress expected scoring to near zero regardless of how dominant the opposing offense is. When Connor Hellebuyck or Igor Shesterkin are “on,” they stop 95-97% of shots — and no feature engineering can fully predict when a goalie will enter that zone.

This is analogous to — but more extreme than — the “dealing” pitcher in baseball. A pitcher having an elite outing might suppress runs to 0-1 over 7 innings. A hot goaltender can suppress goals to 0 over a full 60-minute game against a team generating 35+ shots. The expected goals model might say Team A should score 3.2 goals, but if the opposing goalie posts a .970 save percentage that night, the actual output is 1 goal.

This irreducible variance means NHL models should be calibrated with humility. Maximum edge on any individual bet is lower than in MLB or NBA, but the market compensates by pricing in the variance poorly — especially on props, where books struggle with the same goalie-driven unpredictability.

Shots on goal (SOG) and goaltender saves are high-count stats. A typical NHL team records 28-32 shots per game, meaning individual top-six forwards generate 2-5 SOG per game and starting goaltenders record 24-34 saves. These counts are high enough that a regression model can predict a continuous expected value, then apply a cumulative distribution function (CDF) to compute over/under probabilities at multiple lines simultaneously.

The architecture works as follows: an XGBoost regression model takes 30+ input features and outputs a predicted count (e.g., “Auston Matthews is expected to record 4.2 SOG tonight”). The model also estimates prediction variance. Using a normal CDF centered at the predicted value with the estimated standard deviation, the system calculates the probability of exceeding any threshold — Over 2.5, Over 3.5, Over 4.5, Over 5.5 — in a single pass.

This is the same CDF regression approach used for MLB pitcher strikeouts, where the underlying count (typically 4-10 Ks per game) is high enough to support continuous modeling. The advantage over binary classification is efficiency: one model generates probabilities at every possible line, rather than requiring a separate model for each threshold.

Goals, assists, hits, and points per game are sparse events at the individual player level. The average NHL forward scores roughly 0.25 goals per game. Even elite scorers like Connor McDavid or Auston Matthews average only 0.55-0.65 goals per game. When the average count is below 1.0, regression modeling breaks down — the distribution is heavily zero-inflated, and a normal CDF produces poor probability estimates.

For these stats, binary classifiers are the correct approach. Instead of predicting “how many goals will this player score,” the model predicts “will this player score at least one goal — yes or no?” This maps directly to the most common prop market (anytime goalscorer, Over/Under 0.5 goals) and avoids the mathematical awkwardness of applying a continuous distribution to a nearly binary outcome.

The binary classifier is an XGBoost classification model trained on the same feature set but optimized for log-loss rather than mean squared error. It outputs a probability directly — for example, “Connor McDavid has a 38.2% chance of scoring at least one goal tonight.” If the sportsbook prices his anytime goalscorer prop at +180 (implied 35.7%), the model identifies 2.5 percentage points of edge.

The decision between CDF regression and binary classification follows a practical rule based on the average count of the stat:

The general heuristic: if the average count exceeds approximately 5 per game, CDF regression is viable and preferred because it generates multi-line probabilities from a single model. If the average count is below approximately 1 per game, binary classification is more appropriate. The gray zone between 1 and 5 depends on the specific distribution shape — if it is highly zero-inflated (like hits, where many players record 0 per game), binary is still better even if the average exceeds 1.

Goaltender features are the single most important category in any NHL model. The system tracks: rolling save percentage (last 5, 10, 20 games), goals against average (GAA), save percentage by shot type (wrist, slap, deflection), high-danger save percentage (saves on shots from the slot and crease), home vs. away save percentage splits, performance on back-to-back nights, and days of rest since last start.

The interaction between goaltender quality and opposing team shot volume is critical. A .920 save percentage goalie facing a team that generates 34 shots per game is a fundamentally different prediction context than the same goalie facing a team that generates 26 shots per game. The model captures this through feature crosses — goalie save percentage multiplied by opponent shots per game gives expected goals against, which feeds into totals and prop predictions.

Beyond raw goals for and against, the model uses underlying possession and shot metrics that are more predictive of future performance:

Power play and penalty kill efficiency are among the most impactful features in NHL models. The power play conversion rate across the league averages about 20%, but the spread is massive — elite units like the Edmonton Oilers power play (historically 25-28%) convert nearly 50% more often than bottom-tier units (14-16%).

The model tracks: power play percentage (last 10/20/40 games), penalty kill percentage, power play opportunities per game (some teams draw more penalties through aggressive forechecking), shorthanded goals against rate, and penalty minutes per game for both teams. The interaction between one team's power play and the other team's penalty kill is a critical feature cross — Edmonton's power play against a bottom-5 penalty kill is a fundamentally different prediction than against a top-5 penalty kill.

For player prop models, individual skater features are layered on top of team and goaltender context. The system tracks: rolling shot rate (SOG per 60 minutes), shooting percentage (goals per SOG), time on ice (TOI) — both total and power play minutes, line combination (first line vs. second line, first power play unit vs. second), individual Corsi and expected goals rates, and performance splits by home/away, day/night, and opponent quality.

A critical nuance: line combinations change frequently in hockey. A forward who averages 4.0 SOG per game on the first line with Connor McDavid might average only 2.2 SOG on the second line. The model needs to incorporate current line deployment data, not just season averages. This is especially important for recently traded players or players returning from injury who may slot into different roles.

The remaining features capture game context: rest days (0, 1, 2, or 3+ days since last game), back-to-back indicator (first or second game), travel distance since last game, time zone changes, divisional rivalry indicator (divisional games tend to be lower-scoring and more physical), and season phase (early season, mid-season, playoff push, out-of-contention). Teams eliminated from playoff contention in March show measurably different effort levels, which affects scoring rates and prop outcomes.

The difference between an elite goaltender and a league-average one is numerically small but practically enormous. Connor Hellebuyck at .925 save percentage and a replacement-level goaltender at .900 are separated by just 2.5 percentage points. But on a 30-shot game, that 2.5% gap translates to 0.75 additional goals allowed — nearly half the average winning margin in the NHL.

To put 0.75 goals in context: the average NHL game total is 6.0-6.4 goals. A 0.75-goal swing represents approximately 12% of total scoring. No single player in basketball, football, or baseball has this kind of direct statistical impact on the final score.

One of the most consistent edges in NHL betting comes from backup goaltender starts. When a team's primary goaltender sits and the backup starts, the expected goals against increases by 0.3-0.8 goals depending on the quality gap. Books adjust for this, but the adjustment is often incomplete — especially when the goaltender announcement comes late (morning skate or later) and the book has already set lines based on the expected starter.

The edge is compounded when the backup start is unexpected. If a model is monitoring goaltender news feeds and detects that Igor Shesterkin is sitting for a rest day, it can recalculate all affected predictions (game total, opposing team goals, individual skater goal props for the opposing team) before the book has fully adjusted. This timing advantage — even if it is only 15-30 minutes — is one of the most reliable edges in NHL prop betting.

Additionally, saves props for backup goalies are often mispriced. Books anchor saves lines to the team's average shots against, but they sometimes fail to account for the fact that backup goalies often play behind different defensive deployments (coaches may tighten up their system to protect a weaker goalie) or that the opposing team may adjust their shot volume knowing a backup is in net.

Goaltender performance degrades with workload. A goalie who has started 3 games in 4 nights shows measurably lower save percentages than the same goalie on full rest. The model tracks games started in the last 7, 14, and 30 days, shots faced in recent starts, and days since last start. These workload features help predict not just whether a goalie will start, but how well he will perform if he does start.

This is particularly relevant during the NHL playoff push in March and April, when teams ride their starters harder and backup appearances become less frequent. A goaltender who has started 12 games in the last 20 days is statistically more likely to have an off night than one who has started 8 games in the same span. The saves and goals against predictions adjust accordingly.

A goal in hockey is the product of two sequential events: generating a shot on goal, and that shot beating the goaltender. The first event (shot generation) is primarily a function of the skater's playing style, ice time, and team system — all of which are relatively stable and predictable. The second event (the shot beating the goalie) introduces goaltender-dependent randomness. By betting on shots rather than goals, you are isolating the predictable component and removing the high-variance goaltender filter.

An NHL forward with a 10% shooting percentage who takes 4 shots will score 0.4 expected goals — but on any given night, he might score 0, 1, or occasionally 2 goals. The variance around that 0.4 mean is enormous. The same player's 4 shots have much less variance — he might take 2, 3, 4, 5, or 6 shots, but the distribution is tighter and more normally distributed than the goal-scoring distribution.

The CDF regressor excels at finding value at specific SOG lines. Consider a player the model projects for 3.8 SOG with a standard deviation of 1.6:

In this example, the model finds edge at Over 2.5 and Over 3.5 but not Over 4.5. The ability to evaluate every line simultaneously — rather than just the primary posted line — is a major advantage of the CDF approach. Sometimes the value is at a line the book is not highlighting as the main market, which is precisely where less-sophisticated bettors miss opportunities.

For a deeper explanation of how edge and expected value translate into long-term profitability, see the expected value betting guide.

Not all skaters are created equal for SOG prop betting. The model generates the most confident predictions — and the largest edges — for high-volume shooters on the top line and first power play unit. Players like Auston Matthews (consistently 4.0+ SOG per game), Nathan MacKinnon (3.5+ SOG), and Alex Ovechkin (historically one of the highest shot-volume players in NHL history) provide the deepest data and the most stable shot rates.

The reason is statistical: higher average counts mean lower coefficient of variation (standard deviation divided by mean). A player averaging 4.0 SOG with a standard deviation of 1.6 has a CV of 0.40. A player averaging 1.5 SOG with a standard deviation of 1.2 has a CV of 0.80 — twice as noisy relative to the mean. The model is more confident in its predictions for high-volume shooters, which translates to larger identified edges and better bankroll growth.

Saves props for goaltenders follow a similar logic but at an even higher count. A goaltender projected for 28.5 saves has very tight relative variance, and the CDF model can precisely calculate probabilities at O/U 25.5, 27.5, 29.5, and 31.5. This high-count regime is where the CDF regressor architecture truly excels.

The NFL dominates U.S. sports betting handle, followed by the NBA, then MLB, with the NHL trailing significantly in fourth place. The consequences of this are direct: sportsbooks allocate their sharpest linemakers and most sophisticated models to the sports that generate the most handle. NHL lines, while not sloppy, receive less refinement and less real-time adjustment than NFL or NBA lines.

This shows up empirically in closing line efficiency. Studies of closing line accuracy across sports consistently find that NHL closing lines are less efficient than NFL or NBA closing lines — meaning there is more exploitable gap between the closing line and the true probability. For a model-based bettor, this is the most important structural advantage available.

Sportsbooks set lower maximum bet limits on NHL compared to NFL or NBA. A sharp bettor might be able to place a $10,000 wager on an NFL spread but only $2,000-$3,000 on an NHL moneyline and as little as $500-$1,000 on NHL player props. Lower limits mean less sharp money flows into the market, which means less pressure on books to tighten their lines.

For recreational and mid-level bettors, this is actually good news. The lower limits that prevent whales from fully exploiting inefficiencies do not affect bettors placing $25-$200 wagers. The inefficiency exists at a scale that is inaccessible to the biggest sharps but perfectly accessible to individual bettors using model-driven strategies.

The very thing that makes NHL harder to model — goaltender-driven variance — also scares away the casual betting public. Recreational bettors who get burned by a backup goaltender shutout or a hot goalie stealing a game tend to conclude that “hockey is too random” and migrate their bankroll to sports that feel more predictable (basketball, football).

This behavioral response creates a self-reinforcing cycle: casual bettors leave the market, reducing handle, which reduces book incentive to sharpen lines, which preserves inefficiency, which benefits the model-based bettors who remain. The perceived randomness is real (variance is higher in hockey), but randomness is not the same as unprofitability. A model that correctly estimates probabilities will profit over a large sample regardless of per-game variance — as long as the market is pricing those probabilities inaccurately.

Fewer sportsbooks offer NHL player props compared to NBA or MLB player props. Where you might find 8-10 books posting NBA player points props with tight spreads and competitive odds, you might find only 4-6 books offering NHL SOG props, and only 2-3 offering hits or assists props. Fewer competing market-makers means less price discovery and wider spreads between the best and worst available odds.

This thinness is a double-edged sword: on one hand, the lines are less efficient (good for models). On the other hand, there is less liquidity, meaning you cannot always get the volume you want at the best price. For the individual bettor operating at modest stakes, the liquidity constraint rarely binds — you can almost always get your $50-$200 bet down at the best available line. The inefficiency benefit far outweighs the liquidity cost at this scale.

The NHL regular season is 82 games per team, with 16 teams playing most nights. That is 8-10 games per night, each generating 10-20+ player props. Over a full season, a model evaluating NHL props generates thousands of predictions. If the average edge per qualifying bet is 3-5% (plausible given the market inefficiency described above), the expected return over a season of disciplined, flat-unit betting is substantial.

Compare this to NFL, where a model might generate 50-100 qualified bets over an entire 18-week season. The NHL offers an order of magnitude more volume, with comparable or better per-bet edge. This volume advantage — combined with the structural market inefficiency and goaltender-driven mispricing — makes the NHL arguably the best sport for systematic, model-driven prop betting. It is not the flashiest market, it does not get the most attention, and that is precisely why it works.