Research
The studies behind the model: what the data supports, what it doesn’t, and how the site’s constants were chosen. Each entry links to the full write-up with methods, tables and caveats. Null results are published on purpose — knowing which popular stories the data does not support is as useful as knowing which it does. All numbers are model estimates, not official NHL statistics.
Player evaluation
FindingSeptember 2026
When a skater changes teams, which of his seven rating components come with him and which were the old team’s system? Faceoffs and penalties travel almost fully, even-strength offense mostly, power-play value barely — and a defenseman’s even-strength numbers turn out to be about 80% his teammates’.
6,968 consecutive-season transitions, 1,370 team changes · portability: faceoffs 0.82, penalties 0.82, EV offense 0.68 (F 0.80 / D 0.48), finishing 0.51, EV defense 0.46, PK 0.44, PP 0.26 (D ≈ 0)
READ THE STUDY →Null resultSeptember 2026
Do some players keep producing against elite defenders while others feast on the weak? Elite defensive units really do suppress offense (−16% xGF/60) and elite offense generates even more (+36% xGA/60), but a player’s own gap between his tough and easy minutes does not repeat from one season to the next. “Matchup performer” is noise.
15 seasons of 5v5 segments, opponents tiered by prior-season rating · year-over-year r of the tough-minutes gap: 0.006 (forwards), vs 0.66 for ordinary xGF/60 · only defensemen vs elite offense show a faint career signal (r = 0.25)
READ THE STUDY →Null resultSeptember 2026
Does a sniper + playmaker + grinder line beat three of a kind? No: chemistry lift falls as archetype diversity rises, and the scorer–play-driver–grinder line is the worst composition in the data; two scorers with a play-driver is the best. Lift is measured in expected goals, so a sniper’s finishing is not in it.
967 forward trios, 2,849 duos (5v5, era-adjusted WOWY) · lift by distinct types 1 / 2 / 3: +0.32 / +0.26 / +0.21 xG/60 · GPS line +0.10 vs PSS +0.31
READ THE STUDY →Null resultJuly 2026
Is outperforming in high-importance games a persistent skill? Games are split by playoff-race importance and each player’s realized-over-expected gap in big games is tested for repeatability. It does not repeat. The leaderboard is dominated by goalies, which is start selection by coaches, not clutch.
2021–26, 1,650 players · clutch gap split-half r = 0.047, season-over-season −0.028 · plain per-game value: 0.65 / 0.61
READ THE STUDY →FindingJuly 2026
Sixteen seasons of 5v5 with-or-without-you residuals: together xG% versus what the two players’ apart rates predict. Chemistry exists but is small and noisy; only extreme, high-minute residuals are worth acting on. Bergeron–Marchand post a positive residual in 12 of 13 qualifying seasons.
8,908 qualifying pair-seasons · split-half reliability 0.14 (full-season 0.25) · mean F residual +1.4 xG% points (survivorship)
READ THE STUDY →FindingJuly 2026
Which skills age first? Penalty killing goes earliest and fastest (and defensemen’s never improves). Even-strength defense barely ages after 25. Finishing peaks at 19–24 and slides from the mid-20s; playmaking is flat from 21 to 32. Power-play skill is the most durable value an aging forward has.
Within-player deltas, era-centered, ages 18–38 · F peaks: EV offense 24, PP 27, PK 23, finishing 20, playmaking 28 · 30→35 decline: PK −1.4/yr, EV defense +0.02/yr
READ THE STUDY →MethodJuly 2026
How many games before a rate is believable? Two independent methods — split-half reliability and a walk-forward predictive sweep — agree within a factor of 1.4 on the shrinkage constants the site uses. The old goalie constant was twice too trusting.
skater relWPA ≈ 25–35 pseudo-GP · shooter GAx ≈ 500 SOG · goalie GSAx ≈ 5,000–6,500 shots · all optima are broad plateaus
READ THE STUDY →Roster building & the market
FindingJuly 2026
What does a win cost on the 2026-27 market? Projected wins above replacement for every rostered player, priced against cap hit above the league minimum: the market rate, the best and worst surplus contracts, team totals, what each age band pays versus gets, and the size of the entry-level subsidy.
$5.37M per win above replacement (510 non-ELC contracts) · 640 roster slots · ELCs hold $72.9M of surplus
READ THE STUDY →FindingJuly 2026
Value by draft slot, 2010–2026. The curve is brutally convex at the top with two cliffs (after pick 1 and after pick ~10); below pick ~45 the draft predicts survival, not skill. Top-3 picks are near bust-proof, forwards develop earlier and further than defensemen, and goalie drafting was near-random.
peak wins/82: pick 1 = 6.6, pick 2 = 4.5, pick 3 = 3.6, picks 21–30 = 1.3 · picks 1–3 reach 400 GP 90% of the time, 21–30 34%
READ THE STUDY →FindingJuly 2026
Do coaches give high-leverage ice to the right skaters? Quality is measured on odd games and deployment on even games so the decision is scored against information the coach could have had. Most teams tilt the right way; the wins left on the table are real but modest.
160 team-seasons 2021–26 · mean DeployEff +0.24 (range −0.66 to +0.80) · wins on table 0.49/season, up to 1.3 · r with standings +0.59
READ THE STUDY →MethodJuly 2026
The availability model behind the projections: how much of a season a player can be expected to play, from his recent games-played history and age. Persistence is strong, age costs about a game a year past 27, and a walk-forward bake-off chose the estimator.
16 seasons · year-over-year r of GP-fraction 0.63 · −0.9 GP per year of age past 27 · regulars project to ~0.82 of a schedule
READ THE STUDY →Teams & seasons
FindingJuly 2026
Why did the standings backtest correlation fall from 0.71 to 0.27? Mostly because the league compressed: true-strength spread shrank by 40% while the model’s absolute error improved. Luck is now 41% of points variance, and a perfect strength oracle would top out at r = 0.76. The narrow 2026-27 projection spread is correct behavior.
SD of actual points 18.9 → 13.2; true-strength SD 16.9 → 10.1 · schedule-luck floor 8.5 points · model RMSE 14.0 → 12.8
READ THE STUDY →FindingJuly 2026
What carries from the regular season to the playoffs? The playoff environment tightens only slightly (5v5 xG/60 −3%, all of it shot quality). Even-strength play-driving translates strongly; finishing and grit add nothing once it is in the model. A “playoff performer” effect exists but is barely distinguishable from the regular-season control.
1,395 playoff games, 16 postseasons · r(regular EV offense, playoff xGF/60) = 0.64 · home-ice team wins the series 56%; points + model strength predicts 61%
READ THE STUDY →