Ice hockey is one of the hardest sports to predict. Scoring is low, variance is high, and a single save by the goalie can change the outcome of a game. That is precisely why ProbWin developed an AI system specialized in NHL analysis — one that goes far beyond simple team rankings or win streaks.
Our model integrates advanced statistics (xGF%, Corsi, HDCF%), goaltender performance metrics, and a multi-agent system that waits for starting goalie confirmation before launching any analysis.
Historical result: 65.9% win rate on Moneyline over the last 30 days, climbing to 72.3% on heavy favorites (odds below 1.60).
In this guide, we break down exactly how the system works — from the CatBoost machine learning model to the goalie trigger, advanced metrics, and the multi-agent pipeline that produces every pick.
The CatBoost Model Behind NHL Predictions
Why CatBoost for Hockey?
ProbWin uses CatBoost, a gradient boosting algorithm developed by Yandex, specifically trained on NHL game data. Unlike simpler models (logistic regression, random forests), CatBoost handles the unique challenges of hockey prediction:
- Categorical features: Team matchups, home/away splits, goalie identities — CatBoost handles these natively without one-hot encoding
- Imbalanced data: Hockey favorites win roughly 58% of the time, and the model learns subtle edges within that distribution
- Feature interactions: The interplay between a team's shot quality and the opposing goalie's save ability is captured automatically
- Robustness to noise: Low-scoring games produce noisy outcomes; CatBoost's regularization prevents overfitting to random variance
What the Model Learns
The CatBoost model is trained on thousands of historical NHL games. For each game, it receives a feature vector containing:
Feature categories fed to the CatBoost model:
─────────────────────────────────────────────
Team Offense: xGF, CF%, HDCF%, PP%, goals/game (rolling 10-20 games)
Team Defense: xGA, CA%, HDCA%, PK%, goals against/game
Goalie: Save%, GAA, xGSAA, recent workload, days rest
Situational: Home/away, back-to-back, travel distance
Special Teams: PP% vs opponent PK%, PK% vs opponent PP%
Market: Opening line, current odds, implied probability
The model outputs a win probability for each team. When that probability diverges significantly from what the betting market implies, an edge exists.
Weekly Retraining
The model retrains every Monday at 06:00 UTC using the latest game data. This ensures it captures evolving team dynamics — trades, injuries, hot streaks, coaching changes — rather than relying on stale season-long averages.
Join ProbWin for free
Access our AI predictions for NBA, NHL, Soccer & Tennis. Sign up in 30 seconds.
The Importance of Goalies: The #1 Factor in Hockey
Why the Goalie Changes Everything
In hockey, a goaltender can single-handedly transform a mediocre team into a contender and vice versa. Unlike soccer where the goalkeeper might make 3-5 saves per game, an NHL goalie faces 25 to 35 shots per game. That volume means the difference between a starter and a backup can be worth half a goal or more per game.
Consider this: when a team's starting goalie (Save% .920+) is replaced by a backup (Save% .900), the expected goals against increase by roughly 0.6 to 0.8 per game on 30 shots. That is enough to shift a moneyline by 15-20 cents.
| Goalie Metric | What It Measures | Why It Matters |
|---|---|---|
| Save% | Percentage of shots saved | Overall performance baseline |
| GAA | Goals Against Average | Defensive efficiency per 60 min |
| GSAx | Goals Saved Above Expected | True goalie value vs league average |
| GSAA | Goals Saved Above Average | Cumulative saves above replacement |
| QS% | Quality Start Percentage | Consistency (Save% > .900 in a start) |
GSAx: The Metric That Separates Elite from Average
GSAx (Goals Saved Above Expected) compares a goalie's actual performance to what an average goaltender would have done facing the exact same shots — accounting for shot location, angle, type, and whether it was a rebound or rush.
- Positive GSAx: The goalie stops more than expected — a genuine advantage
- Negative GSAx: The goalie allows more than expected — a measurable handicap
Our model weighs GSAx heavily because it strips away team defense quality. A goalie behind a poor defensive team might have a mediocre Save% but an excellent GSAx, meaning he is actually performing well despite the volume and quality of shots faced. For a deeper analysis, read our full guide on GSAx and elite NHL goalies.
Starter vs. Backup: The Performance Gap
The performance gap between a team's starter and backup is one of the most underappreciated edges in hockey betting:
| Goalie Tier | Typical Save% | Impact on Win Probability |
|---|---|---|
| Elite starter (.925+) | .925 - .935 | +8 to +12% vs league average |
| Good starter (.915-.924) | .915 - .924 | +3 to +7% vs league average |
| Average (.905-.914) | .905 - .914 | Baseline |
| Below average (< .905) | .895 - .904 | -4 to -8% vs league average |
| Emergency backup | .880 - .894 | -10 to -15% vs league average |
When a team's elite starter is scratched and replaced by a backup, the swing can be 10-15 percentage points in win probability. Our model captures this because it receives the actual confirmed goalie's rolling stats, not just the team's season average.
Our Goalie Trigger: No Confirmed Goalie, No Analysis
ProbWin publishes no NHL pick until starting goalies are confirmed. Our system monitors confirmations from DailyFaceoff and official team sources in real time. Only when both goalies for a game are confirmed (or classified as "expected/likely") does our analysis pipeline trigger.
This is a significant advantage: many prediction services publish picks in the morning without knowing who is in net. We wait — because publishing a pick without goalie confirmation is betting blind on a factor that drives 30-40% of the game's outcome.
Goalie Trigger Flow:
────────────────────
DailyFaceoff update → Goalie status check
├─ Both confirmed/likely → TRIGGER analysis pipeline
├─ One missing → WAIT (re-check every cycle)
└─ Backup announced → TRIGGER (with adjusted model input)
Key NHL Metrics: What Our Model Tracks
The Complete Metrics Table
| Metric | Full Name | What It Measures | Predictive Value |
|---|---|---|---|
| CF% | Corsi For % | All shot attempts (on goal + blocked + missed) | Possession dominance |
| FF% | Fenwick For % | Unblocked shot attempts (on goal + missed) | Shot generation quality |
| xGF% | Expected Goals For % | Share of expected goals in team's favor | Overall team strength |
| HDCF% | High Danger Chances For % | Scoring chances from the slot and crease | Offensive quality |
| PDO | Save% + Shooting% | Luck indicator (regresses to 100) | Sustainability filter |
| PP% | Power Play % | Conversion rate on power plays | Special teams offense |
| PK% | Penalty Kill % | Success rate killing penalties | Special teams defense |
xGF% — The Best Single Indicator of Team Quality
xGF% measures the share of expected goals in a team's favor during 5-on-5 play. It is the single most predictive team-level metric in hockey:
- xGF% above 55%: Dominant team — controlling both shot volume and quality
- xGF% 50-55%: Solid team with a positive shot quality differential
- xGF% 45-50%: Struggling team, often relying on goaltending or special teams
- xGF% below 45%: Weak team being outplayed at even strength
Corsi and Fenwick: Possession Metrics
Corsi measures all shot attempts (on goal, blocked, missed). Corsi For% (CF%) indicates which team controls play:
| CF% Range | Interpretation |
|---|---|
| Above 54% | Possession dominance — generating significantly more chances |
| 50-54% | Slight superiority — tilting the ice in their favor |
| 46-50% | Slight disadvantage — opponent controls more of the play |
| Below 46% | Dominated in possession — rarely sustainable for winning |
Fenwick removes blocked shots from the equation, which some analysts prefer since shot-blocking can be a deliberate strategy (teams that block more shots may look worse in Corsi despite playing well defensively).
PDO: The Luck Detector
PDO is simply a team's on-ice Save% plus their on-ice Shooting%. It naturally regresses toward 100 over time. A team with a PDO of 103 is likely running hot and due for regression, while a team at 97 is probably unlucky and due for improvement.
Our model uses PDO as a sustainability filter — it flags situations where a team's recent results may not reflect their true quality.
Special Teams: PP% and PK%
Special teams are a crucial and often decisive factor in NHL outcomes:
- PP% (Power Play): The percentage of power plays converted into goals. The NHL average hovers around 20%. Teams above 25% have elite units; teams below 17% are notably weak.
- PK% (Penalty Kill): The percentage of penalty kills completed without conceding. The NHL average is roughly 80%. A team with PK% above 83% has a structural advantage in close games.
A team with both high PP% and high PK% carries a significant built-in edge, particularly in playoff-style, tightly officiated games.
Two Markets Covered: Moneyline and Totals
Why No Puck Line (Spreads) in Our Model?
Unlike basketball where 5-10 point spreads create meaningful margins, hockey scores are low — typically 2 to 4 goals per team. The standard puck line of 1.5 does not offer enough analytical edge for consistent profitability. A team can dominate every metric and still win 2-1. That is why we focus on two markets where our model has demonstrated a measurable edge:
Moneyline — Who Wins the Game?
Moneyline is the most straightforward market: which team wins the game, overtime and shootout included.
Historical performance snapshot: - 65.9% win rate over a trailing 30-day window - Heavy favorites (odds below 1.60) showed 72.3% win rate historically - The edge has been particularly strong when the model identifies an underpriced favorite — a team whose true win probability exceeds what the odds imply
The confidence window for Moneyline picks is 68 to 85. Picks below 68 lack sufficient edge; picks above 85 typically correspond to odds too short for meaningful returns.
Totals — Over/Under Goals
The Totals market predicts whether the combined goals scored will be above or below a given line (usually 5.5 or 6.0 in the NHL).
| Factor | Typical Impact on Totals |
|---|---|
| Two elite goalies | Under — fewer goals expected |
| Two high-tempo, weak-defense teams | Over — run-and-gun style |
| Back-to-back for one team | Unpredictable — fatigue can cause blowout or tight checking |
| Rivalry / playoff atmosphere | Under — tight, physical games |
| Backup goalie on one side | Over — weaker goaltending inflates expected goals |
| Power play matchup (elite PP vs weak PK) | Over — special teams generate extra scoring |
Totals analysis is inherently more complex than Moneyline because it requires evaluating both teams' offensive output AND both goalies simultaneously. Our model examines the interaction between team shot quality and goalie save ability to project a total goals expectation.
Concrete Example: How a Pick Gets Made
Here is a walkthrough of how our system analyzes a hypothetical game, illustrating each step:
Game: Rangers at Canadiens
Step 1 — Data Collection (Scout)
Rangers (rolling 15 games):
xGF%: 56.2% | CF%: 53.1% | HDCF%: 54.8%
PP%: 24.3% | PK%: 82.1% | Goals/Game: 3.41
Canadiens (rolling 15 games):
xGF%: 47.8% | CF%: 48.2% | HDCF%: 46.1%
PP%: 18.1% | PK%: 78.4% | Goals/Game: 2.67
Confirmed Goalies:
Rangers: Shesterkin — Save% .928, GAA 2.12, GSAx +8.4
Canadiens: Montembeault — Save% .905, GAA 3.18, GSAx -2.1
Step 2 — Model Prediction
The CatBoost model processes these features and outputs: - Rangers win probability: 68% (implied fair odds: ~1.47) - Canadiens win probability: 32%
Step 3 — Edge Detection
The bookmaker posts Rangers at 1.58 (implied probability: 63.3%).
Edge = Model probability - Implied probability = 68% - 63.3% = +4.7%
This exceeds the minimum edge threshold. Confidence score: 72 (within the 68-85 window).
Step 4 — Executor Validation
The Executor confirms: - Confidence 72 is within [68, 85] range - Both goalies are confirmed - No conflicting pick exists on this game - Odds are above minimum threshold
Result: Rangers ML pick is issued.
The NHL Multi-Agent Pipeline
Our NHL system operates with the same proven five-agent architecture used across all sports, adapted with hockey-specific logic:
1. Scout — Hockey-Specialized Data Collection
The NHL Scout gathers and structures all relevant data: - Rolling team stats over 10 and 20 games (xGF%, CF%, HDCF%, PP%, PK%) - Confirmed goalie stats (Save%, GAA, GSAx, recent workload, days of rest) - Current Moneyline and Totals odds from multiple sources - MoneyPuck advanced metrics for shot quality data - Head-to-head history and recent results
The Scout uses a 300-second timeout with two retries, ensuring robust data collection even when external APIs are slow.
2. Analyst — Contextual Game Evaluation
The Analyst is powered by Claude Opus and evaluates each game by integrating: - Relative team strength based on underlying metrics (not just record) - Goaltender advantage or disadvantage (the most critical factor) - Fatigue and schedule factors (back-to-backs, travel, time zone changes) - Recent performance trends with context (a 5-game losing streak means less if the underlying xGF% is still strong) - Special teams matchup asymmetries
The Analyst produces a structured decision: JOUER (play the model pick), INVERSER (reverse direction), CREER (create a new pick the model missed), or PASSER (skip the game entirely).
3. Executor — Rigorous Validation
The Executor is pure Python logic (no AI) that enforces hard rules: - Confidence must fall within the 68-85 window - Available odds must exceed the minimum threshold - No conflicting picks on the same game (cannot have both ML sides) - All required data must be present
4. Writer — Contextualized Analysis
The Writer produces a detailed game analysis mentioning both confirmed goalies, key statistical matchups, and the reasoning behind the pick. Every analysis includes a mandatory "Confirmed Goalies" section.
5. Learner — Continuous Improvement
The NHL Learner runs every morning after games are settled. It: - Reviews each settled pick (win or loss) with goalie-specific analysis - Tracks performance by confidence bin (60-64, 65-69, 70-74, 75+) - Monitors goalie tier performance (elite, good, average, poor) - Identifies patterns by market, side, and odds range - Can adjust confidence thresholds by up to 3 points per run (Mondays only) - Updates the playbook with fresh insights for the next analysis cycle
Why Goalie Confirmation Is Our Differentiator
How the Trigger System Works
- Continuous monitoring: Our system checks goalie announcements via DailyFaceoff and official team sources throughout the day
- Confirmation: When both goalies for a game are confirmed (or classified as "expected/likely"), a trigger fires automatically
- Pipeline launch: The Scout, Analyst, Executor, and Writer pipeline starts — typically 2-4 hours before game time
- Publication: Picks are published as soon as the full analysis is complete
The Cost of Ignoring Goalie Status
A goalie change can shift moneyline odds from 1.60 to 2.00 or beyond. Here is why:
| Scenario | Impact |
|---|---|
| Elite starter confirmed | Odds tighten — market prices in the advantage |
| Backup announced last-minute | Odds drift 10-20 cents — value shifts dramatically |
| "Expected" starter scratched | Line moves sharply — pre-confirmation picks become stale |
| Both teams start backups | Total typically moves up 0.5 goals |
Publishing a pick without knowing the goalie is betting blind on a factor that represents the single largest variable in game outcome. Our system eliminates this risk entirely.
How to Use NHL Picks Effectively
Bankroll Management
- Flat betting: Allocate 1-3% of your bankroll per pick. Hockey variance demands discipline.
- Patience: Hockey produces more variance than basketball due to low scoring. Losing streaks of 4-5 picks happen even with a 65%+ long-term win rate.
- Follow the system: Do not cherry-pick only the selections that "look good" to you. The edge comes from volume and consistency.
- Moneyline priority: Our historical edge has been stronger on ML than on Totals.
For a comprehensive guide on sizing bets and managing drawdowns, see our bankroll management guide.
Hockey-Specific Tips
- Overtime always counts: In NHL regular season, there is always a winner (overtime + shootout). Our ML picks include overtime.
- High variance is normal: A game can turn on a single power play or a fluky bounce. It is expected for 70% favorites to lose regularly.
- Long season creates opportunity: 82 games per team means large sample sizes and consistent edge opportunities over weeks and months. Think in terms of 50+ pick samples, not individual results.
What to Avoid
- Do not bet on games our system has passed — those are passed for a reason
- Do not stack multiple NHL picks into parlays — the variance compounds
- Do not chase streaks in either direction ("this team won 5 in a row" is not a signal the model missed)
- Do not ignore last-minute goalie changes — if the confirmed goalie is scratched after our pick is published, the analysis may no longer apply
Frequently Asked Questions
How accurate are AI predictions for NHL games?
No prediction system is accurate on every individual game — hockey's inherent variance makes that impossible. What matters is the long-term edge. Our model has historically achieved a 65.9% win rate on Moneyline over 30-day rolling windows, meaning it identifies value the market misses roughly two-thirds of the time. The edge is statistical, not absolute.
Why does the model not cover the puck line (1.5 spread)?
The puck line at -1.5 or +1.5 is essentially an all-or-nothing proposition in a low-scoring sport. Most NHL games are decided by 1-2 goals, making the 1.5 spread extremely binary. Our model's edge is in identifying which team wins, not by how much. Moneyline and Totals provide more consistent value because they align with what our features actually predict: win probability and total goals expected.
What happens when a goalie is changed after the pick is published?
If a confirmed goalie is scratched after our analysis is published, the pick's underlying thesis may no longer hold. We recommend treating the pick with caution or skipping it entirely. The goalie is often the single largest factor in our model's output, so a change in net changes the math.
How does the model handle back-to-back games?
Back-to-back games are a feature in the model. The system accounts for fatigue, travel distance, and which goalie is likely to start (teams almost always use their backup in one of the two games). The goalie trigger ensures we know exactly who is in net before making any prediction, which is especially critical in back-to-back situations.
Can I combine NHL picks with other sports?
We recommend flat betting each pick independently. Cross-sport parlays compound variance and eliminate the mathematical edge that makes individual picks profitable over time. If our NHL pick has a 65% probability and our NBA pick has a 70% probability, combining them drops to a 45.5% expected hit rate — below breakeven at standard parlay odds. For more on bet types, see our guide on Moneyline, Spread, and Totals.
View Today's NHL Picks
Our AI analyzes every NHL game once goalies are confirmed. The full pipeline — from data collection to published analysis — runs automatically, ensuring every pick reflects the most current information available.
Explore today's predictions on our NHL picks page.
Deepen your hockey analytics knowledge with our guides on Corsi and Fenwick possession metrics and GSAx: the metric that reveals elite goalies.
