Colorado destroying St. Louis 6-1. Minnesota shutting out Boston 4-1. New Jersey holding off Philadelphia 3-2. The first week of the 2026-27 NHL season delivered its share of upsets and confirmations. Our AI model tracked every game — here's what it read, and what it tells us about the season ahead.
The 2026-27 regular season opened September 29 on an 84-game schedule, the longest in NHL history under the new CBA. On opening night, reigning Stanley Cup champion Carolina hosted Florida in a clash of the last three champions.
Key takeaways - Our AI issued 10 official ML picks over the week of October 1–7: 6 wins, 4 losses (60%, n=10 picks) - Over the past 30 days, the model runs at 63.6% win rate on 11 ML picks and 61.5% on 13 totals picks (+2.29 units net) - Three patterns stand out from opening week: goaltending is decisive, B2B fatigue matters by day 3, and OVER lines are overpriced early in the season - Ten games tonight (October 8) including CGY–COL and CAR–VAN — analysis below
Opening Week by the Numbers
From October 1 through 7, our model covered the following matchups. Scores are real; probabilities come directly from the ProbWin ML model:
| Game | Date | Final Score | Our Pick | Model Probability | Result |
|---|---|---|---|---|---|
| New Jersey – Philadelphia | Oct 1 | 3-2 | HOME (Devils) | 76.6% | ✅ WIN |
| Carolina – Washington | Oct 2 | 2-5 | AWAY (Capitals) | 63.3% | ✅ WIN |
| Buffalo – Chicago | Oct 3 | 4-3 | HOME (Sabres) | 73.7% | ✅ WIN |
| Tampa Bay – Washington | Oct 3 | 3-1 | HOME (Lightning) | 69.7% | ✅ WIN |
| Pittsburgh – Montreal | Oct 3 | 6-5 | AWAY (Canadiens) | 74.0% | ❌ LOSE |
| Minnesota – Boston | Oct 3 | 4-1 | HOME (Wild) | 74.0% | ✅ WIN |
| Colorado – St. Louis | Oct 3 | 6-1 | HOME (Avalanche) | 80.2% | ✅ WIN |
| San Jose – Los Angeles | Oct 3 | 5-4 | AWAY (Kings) | 71.0% | ❌ LOSE |
| Boston – Ottawa | Oct 5 | 1-4 | HOME (Bruins) | 54.7% | ❌ LOSE |
| Chicago – St. Louis | Oct 6 | 4-2 | AWAY (Blues) | 63.1% | ❌ LOSE |
ML record: 6W/4L on 10 settled picks. The 4 losses share a common thread: either a weak model edge (Boston–Ottawa at 54.7%, Chicago–St. Louis at 63.1%) or an unpredictable upset (San Jose beating LA 5-4 despite clear xGoals inferiority).
On the totals side, 9 official picks this week — 5W/4L (55.6%):
| Game | Goals | Our Pick | Line | Model Prediction | Result |
|---|---|---|---|---|---|
| San Jose – Florida | 7 | OVER 6.5 | 6.5 | 6.69 | ✅ WIN |
| Detroit – NY Rangers | 2 | OVER 6.0 | 6.0 | 6.75 | ❌ LOSE |
| Carolina – Washington | 7 | OVER 6.0 | 6.0 | 6.77 | ✅ WIN |
| Seattle – Calgary | 7 | OVER 5.5 | 5.5 | 6.75 | ✅ WIN |
| Boston – Ottawa | 5 | UNDER 6.0 | 6.0 | 5.40 | ✅ WIN |
| Detroit – Ottawa | 8 | OVER 6.0 | 6.0 | 6.62 | ✅ WIN |
| Buffalo – Minnesota | 5 | OVER 6.5 | 6.5 | 7.05 | ❌ LOSE |
| Los Angeles – Florida | 3 | OVER 6.0 | 6.0 | 6.49 | ❌ LOSE |
| Buffalo – Chicago | 7 | UNDER 6.5 | 6.5 | 5.62 | ❌ LOSE |
Join ProbWin for free
Access our AI predictions for NBA, NHL, Soccer & Tennis. Sign up in 30 seconds.
5 Games Our AI Read Correctly — and What It Saw
1. Colorado 6-1 St. Louis (October 3) — Model Pick: HOME at 80.2%
The Cup favorite played like a Cup favorite. Colorado finished last season at 121 points (Presidents' Trophy), and the model gave them 80.2% probability. The 6-1 final confirmed the gap in class.
What the model calculated: a massive xGoals differential over 10 weighted games, excellent road form for Colorado, and a St. Louis goaltender with a weak recent form rating. Model edge over market odds: 9.2 percentage points.
2. Minnesota 4-1 Boston (October 3) — Model Pick: HOME at 74.0%
Minnesota started the season as the #1 team in the power rankings (DailyFaceoff, October 5). But at pick time, the market only priced them at 67% — that gap was the edge. The Wild validated it with a dominant 4-1 performance.
Boston was playing on a back-to-back with an incomplete defensive corps. Our model treats B2B fatigue as a primary factor — that's what tilted the balance toward Minnesota. A pattern documented across 7 seasons of walk-forward testing.
3. Washington 5-2 Carolina (October 2) — Model Pick: AWAY at 63.3%
The reigning champion loses at home on night two of the season. Our model had flagged a Washington goaltender in strong form (20.5 percentage points of edge). No sentiment, no defending-champion bias: the data spoke, and Washington won 5-2 in Raleigh.
4. Boston 1-4 Ottawa (October 5) — UNDER 6.0, model prediction 5.40 goals
The goaltender duel: Jeremy Swayman (BOS) vs. Linus Ullmark (OTT). Ullmark knows the Bruins' tendencies inside out — he played there for several seasons. Final: 5 goals, UNDER validated, model prediction of 5.40 confirmed. The edge on this pick was 0.60 goals.
This is exactly the type of matchup the model loves: two goalies in form, defensive styles on both sides, and a 6.0 line the market had overpriced.
5. New Jersey 3-2 Philadelphia (October 1) — ML HOME at 76.6%
Opening week, tight game. The Devils at 76.6% confidence — high for an early-season game. The model had factored in New Jersey's strong home-ice advantage and early-season form against a Philadelphia team missing several key pieces. Final score 3-2, pick validated.
To understand how our model builds these probabilities game by game, read the complete NHL AI model guide.
What Our Model Actually Looks At
Every NHL game goes through four filters before a pick is issued:
1. Goaltender form (last 7 games, weighted)
→ Goalie Rating T3: save % vs. league average, recent trend
→ A hot goalie can flip a matchup that looks one-sided on paper
2. Defensive xGoals matchup
→ Expected goals for and against, adjusted by team style
→ Corsi + Fenwick over a 10-game rolling window
3. B2B fatigue
→ Calibrated penalty for teams playing their 2nd game in 24 hours
→ Pattern S1 (spread 4-8, home team on B2B → AWAY): 65.4% WR historically
4. Line shopping (8 books compared)
→ The recorded odds = best available at pick time
→ Actual edge, not inflated by an underpriced average line
The model does not vote. It estimates a probability, compares it to market odds, and only issues a pick when the gap clears the confidence threshold. No narrative, no gut feeling.
30-day performance (as of October 8, 2026):
| Market | Picks (n) | Win Rate | Net PnL |
|---|---|---|---|
| ML (winner) | 11 picks | 63.6% | +0.34 u |
| Totals (OVER/UNDER) | 13 picks | 61.5% | +2.29 u |
A sample of 11-13 picks is too small to claim a statistically proven edge. Hockey variance is structurally higher than in other sports. What matters is consistency over 200+ annual picks.
For a deeper dive on NHL totals analysis: NHL Over/Under Betting Guide — 6 Key Factors.
Tonight's Games (October 8) — What the Model Will Watch
Ten games on the slate. Starting goaltenders are not yet all confirmed — which is exactly why the model waits before issuing official picks.
| Game | Key analytical angle |
|---|---|
| CAR vs. Vancouver | Reigning champ: CAR goaltender is decisive; VAN has strong offensive upside |
| Calgary vs. Colorado | Colorado (Cup favorite, 121 pts last season) away in Calgary |
| Boston vs. Utah | Boston looking for first win after back-to-back losses |
| Tampa Bay vs. Minnesota | Style clash: Minnesota's hot start vs. Tampa's veteran depth |
| Ottawa vs. Philadelphia | Ottawa 3W in opening week — real momentum or small-sample noise? |
CGY–Colorado is analytically the most interesting. Colorado has a potentially high model edge (depth, form, Stanley Cup favorite per NHL.com staff poll) but plays on the road. That is exactly the tension the model must resolve: high confidence vs. road constraint.
Tonight's official picks go live on the NHL dashboard once goalies are confirmed (typically 3-4 hours before puck drop).
Check out our complete NHL 2026-27 season guide for background context on every team.
Frequently Asked Questions
Is the ProbWin model reliable over a full NHL season? The model has run in live conditions since the 2025-26 season. A 10-pick opening week sample is statistically too short to draw conclusions — wait for 50+ picks before reading results seriously. What matters: methodological consistency, not week-1 noise.
Why do some high-confidence OVER picks lose? BUF–MIN (predicted 7.05 goals, final score 5) is a classic example. The model correctly estimated offensive potential — but two hot goaltenders can shut down any total. Hockey variance is high by design. Over 50+ totals picks, these misses smooth out.
What's the difference between an "official pick" and an "AI trend"? An official pick passes the confidence threshold AND goaltender confirmation. An AI trend means the model has an opinion, but the edge or goalie confirmation is insufficient to validate. Only official picks enter our performance record.
When are NHL picks available each day? The pipeline runs around 9 AM UTC. Picks are finalized once starting goalies are confirmed, typically 3-4 hours before game time. Subscribers receive an instant Telegram notification.
Are the odds you use the best available? Yes. ProbWin compares odds from 8 books in real time. The recorded odds = the best available at pick time. This maximizes real edge and ensures our value calculations are honest — not based on an underpriced average line.
Next Step
Opening week confirmed several core hypotheses: Colorado and Minnesota are as strong as expected; goaltenders continue to define tight games; and UNDER bets can surprise early in the season when teams have not yet found their offensive rhythm.
To go further: - Understand the model in depth → NHL AI Predictions: How Our Model Works - Full 2026-27 season context → NHL 2026-27 Season Guide - Analyze NHL totals yourself → NHL Over/Under Betting Guide - Player props (shots on goal) → NHL Player Props: Shots on Goal
