NHL Playoff Prediction Methodology

How we compute series win probabilities for the 2026 Stanley Cup Playoffs.

Overview

Our model predicts the probability each team wins a best-of-seven playoff series using four factors combined into a per-game win probability, then run through a 10,000-simulation Monte Carlo of the remaining games. Predictions update live after every game played and automatically incorporate in-series data once three games have been played.

The weights for each factor were derived empirically from five seasons of historical playoff data (2020-21 through 2024-25, 75 series total) — not from intuition. We measured each factor's correlation with actual series outcomes and assigned weight proportional to predictive power.

The four factors

Elo rating (37% weight)

Elo is a continuously-updated team strength rating that incorporates every game result, opponent quality, and margin of victory. It was the single strongest predictor of playoff series outcomes in our historical analysis. Our Elo implementation updates after every game, so by the start of playoffs each team's rating reflects their full regular-season body of work — including late-season form.

Note: Elo correlates strongly with season goal differential (r=0.76), points percentage (r=0.76), and 5v5 xGF% (r=0.67). Including all of those separately would be triple-counting team strength. Using Elo alone captures the combined signal cleanly.

Starting goaltender GSAx/60 (28% weight)

Goals Saved Above Expected per 60 minutes. A shot-quality-adjusted measure of goaltender performance: positive numbers mean the goalie is stopping more pucks than an average NHL goalie would given the quality of chances faced.

GSAx is nearly independent from Elo (r=0.014 in our analysis) — a team's goaltender quality isn't baked into their team strength rating, making this a genuinely additive factor. We use current-season data only; historical multi-year goalie data has poor year-to-year repeatability.

Why not raw save percentage (SV%)? In our analysis, raw SV% had essentially zero predictive power (r=0.012 all rounds, -0.117 for Round 1). It's a well-known but noisy stat because it doesn't adjust for shot difficulty. GSAx is the correct metric.

Power play xGF per GP (20% weight)

Expected goals generated on the power play, per team game played across the season. We use expected goals rather than raw goals because xG captures chance quality — a team that creates high-danger scoring chances on the power play will continue to convert in the playoffs even if their regular-season finishing was unlucky.

Penalty kill GA per GP (15% weight, inverse)

Goals allowed on the penalty kill per team game played. Lower is better, so this factor favors the team with the stronger penalty kill. In tight playoff games, preventing power play goals is a meaningful differentiator.

How the weights were determined

We ran a correlation analysis on 75 playoff series (2020-21 through 2024-25) — 40 Round 1 series plus 35 later-round series. For each potential factor, we computed:

• The differential between the two teams in that factor (higher seed minus lower seed)
• Pearson correlation between that differential and the series outcome (higher seed won = 1, lost = 0)

Factors with higher correlations received proportionally more weight. We also checked pairwise correlations between factors to avoid including multiple metrics that measure the same underlying thing — hence using Elo alone as our team-strength factor rather than stacking goal differential, points %, and CF% on top.

Factors we considered and dropped

Several metrics that are frequently cited in playoff predictions didn't hold up against our data:

• Raw save percentage: r=0.012 overall, slightly negative for Round 1. Noise.
• Recent form (last 20 games goal differential): r=-0.092 in Round 1 — slightly negative. Teams peaking in the final weeks of the regular season tend to underperform in playoffs, possibly because late-season games are often meaningless for playoff-bound teams and rotation schedules are disrupted.
• Head-to-head regular-season record: Small-sample noise. Teams play 3-4 regular-season games against each other; it's not enough data to be predictive once team-strength factors are already accounted for.
• Blocked shots, hits: Often treated as "grit" indicators in traditional analysis but had near-zero correlation. These are often consequencesof playing defensively, not causes of success.
• Individual player points: Weak signal. We tested whether the top scorer's regular-season production predicts playoff advancement — it doesn't, at least not once team-level factors are included.

Per-game win probability

The four factors plus home ice are combined into a single per-game win probability for the higher-seeded team:

per_game_prob = 0.50
  + elo_shift        // max ±0.08, scaled by 0.04 per 100 Elo points
  + gsax_shift       // max ±0.06, scaled by 0.03 per 0.5 GSAx/60
  + pp_shift         // max ±0.04, scaled by 0.02 per 0.25 PP xGF/GP
  + pk_shift         // max ±0.03, scaled by 0.015 per 0.25 PK GA/GP
  + home_ice_shift   // ±0.04 depending on venue

Clamped to [0.28, 0.72]

The clamp exists because extreme factor alignment in a small sample can produce mathematically unrealistic per-game probabilities. In practice, the best historical Round 1 favorites produced series win probabilities around 75-78%.

Monte Carlo simulation

Once we have per-game probabilities for each of the remaining games in the series (games already played are locked in at their actual result), we run a 10,000-iteration Monte Carlo simulation. In each iteration:

1. Start from the current series score (e.g., 2-1 after Game 3)
2. Simulate the next scheduled game using its venue-specific per-game probability
3. Increment the winner's series score
4. Continue until one team reaches 4 wins, or we exhaust the remaining schedule

The final series win probability is the fraction of simulations where each team reached 4 wins first. We also track series length distributions (4-, 5-, 6-, 7-game outcomes) to produce the "most likely outcome" shown on each series page.

Live updates during a series

Predictions recompute on every page load, automatically reflecting every game that's been played:

• Elo updates after every game via our existing rating cron.
• Goaltending (GSAx/60), power play, and penalty kill switch from season data to in-series data once three games have been played in the series. This gives us enough sample to meaningfully represent how each team is actually performing in this matchup, not the regular season.
• The Monte Carlo re-simulates only the remaining games starting from the current series score — so a 2-0 lead naturally produces a dramatically higher series win probability even before any factor adjustments.

Validation against historical series-state base rates

We verified that our Monte Carlo produces realistic probabilities at each possible series state by comparing against historical NHL data:

These match within 1-6 percentage points. The small empirical "extras" likely reflect real momentum or lineup adjustments that aren't in our model, but the gap is small enough that we don't apply an additional correction.

Round-specific behavior (and a caveat)

Historical higher-seed series win rates by round:

• Round 1: 70%
• Round 2: 35%
• Round 3 (Conference Final): 30%
• Stanley Cup Final: 80% (small sample: 5 series)

By Round 2 and beyond, the remaining teams are much more evenly matched in true talent regardless of regular-season seeding — upsets are common and "being the higher seed" stops meaning much.

Current weights are calibrated on Round 1 data. When Round 2 begins, we'll recompute weights specifically from Round 2+ historical data, which will likely reduce the impact of seeding-driven factors (Elo, goal differential) relative to goaltending and special teams.

Known limitations

• No injury adjustment. We experimented with an injury factor but dropped it because the implementation was too rough — we couldn't reliably weight injuries by player importance. A starting goalie being out is hugely material and currently isn't reflected. Use judgment when a key player is known to be out.
• Flat home-ice advantage. We use a single ±4% per-game shift for home ice across all teams. Some teams historically have stronger home-ice advantages than others, but per-team playoff home game samples are too small to calibrate reliably.
• Five seasons of training data. Our weights come from 2020-21 through 2024-25. That's 75 series — enough to identify strong signals but limited statistical power for finer-grained effects.
• Monte Carlo variance. With 10,000 simulations, reported probabilities have approximately ±0.5 percentage point noise. A prediction that changes from 54% to 55% between page loads is likely just simulation variance, not a real model shift.

Where to see predictions on the site

• The full bracket shows each team's current win probability next to their win count.
• The home page shows first-round matchups with win probabilities.
• Individual series pages (linked from the bracket) show the full factor breakdown, pre-series vs current prediction swing, and historical series-state context.

Questions or feedback? Model tuning is ongoing — we'll refine these weights after each playoff year as more data accumulates.