AI NHL Predictions: How Machine Learning Analyzes Hockey

Hockey is one of the most unpredictable major sports. A single hot goaltender can steal a game against a far superior opponent. A power play goal in the third period can flip the outcome. An 82-game regular season produces enough variance that even the best teams lose 25+ games. This makes NHL prediction both challenging and valuable — when a model can identify patterns that the market misses.

Our NHL prediction model was built specifically for hockey's unique characteristics. It doesn't treat hockey like football with ice. Every feature, every weighting, and every data pipeline is designed around what actually determines NHL game outcomes. Here's how it works.

The Single Most Important Factor: Goaltending

In no other major sport does one player's performance dominate the outcome as much as the starting goaltender in hockey. A goalie having a .940 save percentage night can single-handedly win a game for a mediocre team. A goalie having a .880 night can sink a Cup contender.

Our model tracks goaltender performance across multiple dimensions:

• Save percentage over the last 10, 20, and full-season games. Recent form is weighted more heavily because goaltender performance fluctuates significantly throughout the season.
• Home vs away save percentage. Most goalies perform measurably better at home. The gap is typically 5-15 points of save percentage (.920 home vs .910 away, for example).
• Performance against high-danger scoring chances. Not all shots are equal. A goalie who excels at stopping high-danger chances (from the slot and crease area) is more valuable than one with a high save percentage driven by easy perimeter saves.
• Workload and rest. A goaltender who started the previous night faces measurable performance degradation. Our model tracks days since last start and total starts in the last 7 and 14 days.
• Matchup history. Some goalies historically struggle against specific teams. Whether it's a stylistic mismatch or just random variance, the data is incorporated.

Goaltender confirmation (knowing who is starting) is critical. Our model updates predictions once starting goalies are confirmed, which typically happens the morning of the game. An unconfirmed goalie matchup adds uncertainty that the model explicitly accounts for.

Team-Level Features

Beyond goaltending, our model analyzes a comprehensive set of team-level statistics.

Expected Goals (xG)

Expected goals is the most important team-level metric in modern hockey analytics. It measures the quality of shots a team generates (offensively) and allows (defensively), based on shot location, shot type, and game situation. A team that generates 3.5 xG per game but only scores 2.8 goals is likely due for positive regression — their shot quality is better than their results suggest.

Our model uses both raw xG and the xG differential (offensive xG minus defensive xG) as primary features. Teams with strong xG differentials tend to outperform their records over time, and the model catches this divergence early.

Power Play and Penalty Kill

Special teams performance varies dramatically across the NHL. The difference between the best power play (25%+) and the worst (12-15%) is enormous. Our model tracks:

• Power play percentage (recent and season-long)
• Penalty kill percentage (recent and season-long)
• Penalties drawn per game for each team (some teams create more power play opportunities than others)
• Penalties taken per game (undisciplined teams give up more short-handed situations)

When a team with a 28% power play faces a team with a 72% penalty kill, the model quantifies the expected special teams advantage and adjusts the win probability accordingly.

Faceoff Win Percentage

Faceoffs directly impact possession. A team that wins 54% of faceoffs starts more possessions with the puck, which compounds over 60+ faceoffs per game. Our model tracks team-level faceoff percentages and, when available, the specific faceoff matchups between starting centers.

Schedule and Fatigue Factors

The NHL regular season is a grind. 82 games in roughly 180 days, with extensive travel. Fatigue is real, measurable, and often underpriced by the betting market.

Back-to-Back Games

This is the most exploitable scheduling factor in NHL prediction. Teams playing the second game of a back-to-back set perform measurably worse — historically, about 3-5% lower win rate compared to well-rested opponents. The effect is even stronger when:

• The team traveled overnight between cities
• The previous game went to overtime
• The backup goaltender starts (which happens roughly 60% of the time on the second night)

Our model quantifies each of these sub-factors individually. A back-to-back at home with the starter is very different from a back-to-back on the road with the backup after an OT loss.

Three Games in Four Nights

Less discussed than back-to-backs but still impactful. When a team plays three games in four nights, performance tends to decline in the third game, particularly if travel was involved. The model tracks this specific scheduling pattern.

Travel Distance and Time Zones

A team flying from Vancouver to Miami for a road game is at a different disadvantage than one driving from Philadelphia to New York. Our model calculates travel distance between cities and applies fatigue adjustments based on both distance and time zone changes. Cross-country trips (especially west-to-east) carry the largest penalty.

Momentum and Recent Form

Hockey teams go through hot and cold stretches that affect near-term performance. Our model tracks recent results with exponential time decay — games from last week matter more than games from last month.

But we go beyond simple win/loss streaks. The model evaluates the quality of recent performance:

• Were the wins against playoff-caliber teams or bottom-feeders?
• Were the losses close (one-goal games, overtime) or blowouts?
• Is the team's shooting percentage abnormally high or low (suggesting regression)?
• Is the goaltender on a hot streak or cold streak?

A team that's 7-3 in their last 10 but all 7 wins came against bottom-10 teams is different from a 7-3 team that beat three division leaders. The model distinguishes between these situations.

How Predictions Are Generated

All of these features — goaltender data, team stats, special teams, scheduling, travel, and recent form — are fed into a gradient-boosted ensemble model. The model outputs a win probability for each team in every game.

Predictions are published daily on the NHL Picks page. Each prediction includes the matchup, the predicted winner, the win probability, and the key factors driving the prediction. When starting goalies are confirmed, predictions are updated to reflect the specific goaltender matchup.

Where the Model Finds Edges

The NHL betting market is less efficient than the NFL market due to lower betting volume and less public attention on many games. This creates more opportunities for the model to find value.

The biggest edges typically appear in:

• Back-to-back situations where the market doesn't fully adjust for fatigue and backup goaltenders
• Early-season games when the market is still relying on prior-season ratings and the model has already incorporated current-season data
• Goaltender switches announced close to game time, before the market line fully adjusts
• Regression spots where a team's PDO (shooting percentage + save percentage) is unsustainably high or low

Model Performance

Our NHL model's accuracy is tracked in real time on the dashboard. Over meaningful sample sizes, the model achieves moneyline accuracy in the 57-62% range for predicted favorites, with the strongest performance in games involving clear fatigue differentials or goaltender mismatches.

We are transparent about where the model struggles. Overtime games are essentially coin flips once regulation ends. Playoff series get harder to predict after Game 1 because coaches make adjustments. And hot goaltender streaks can override even the strongest model signals.

The Bottom Line

NHL prediction requires a model that understands what makes hockey different: the dominance of goaltending, the impact of schedule fatigue, the volatility of special teams, and the pace of a sport where a single shift can change the outcome.

Our model was built from the ground up for hockey. It's not a generic "all sports" model with hockey data thrown in. Every feature is hockey-specific. Every weighting reflects how hockey games are actually won and lost.

Check today's NHL picks to see the model in action. Check the dashboard to see its historical track record. And if you want the full picture of how we handle all five sports, the entire platform costs 99 cents for lifetime access.