Run 2026-07-08 on the 2021-26 rated-games universe. Design: per player-season, games split by per-season terciles of w_game (playoff-race importance); clutch_gap = per-game mean of (wpa - xgwpa) in HIGH-importance games minus the same in LOW-importance games. Persistence tested split-half within season (odd/even games, >=30 GP per half) and season-over-season (>=30 GP both seasons), against a 200-rep shuffled-label null. Positive control: plain per-game wpa. Produced by clutch_study.py.
======================================================================== CLUTCH STUDY -- is outperforming under pressure a persistent skill? ======================================================================== Ledger rows 236,268 | games 6,198 | seasons 2021-2025 | players 1,650 Games in HIGH/LOW importance terciles: 4,135 (middle tercile dropped) league mean resid: LOW -0.003 HIGH -0.021 (pp per player-game) league mean xgwpa: LOW +0.465 HIGH +0.481 (pp per player-game) league mean wpa: LOW +0.462 HIGH +0.460 (pp per player-game) Eligible player-seasons: split-half (>= 30 GP/half) 2,251; season-pairs (>= 30 GP both years) 2,275 ------------------------------------------------------------------------ 1. PERSISTENCE OF THE IMPORTANCE GAP (HIGH-bucket mean - LOW-bucket mean) ------------------------------------------------------------------------ metric split-half r n season/season r n clutch_gap (wpa-xgwpa) +0.047 1912 -0.028 1749 xgwpa gap (process) +0.079 1912 +0.054 1749 wpa gap (results) +0.055 1912 +0.031 1749 Robustness -- same test, halves = early/late season instead of odd/even metric split-half r n clutch_gap (wpa-xgwpa) -0.109 760 xgwpa gap (process) -0.002 760 wpa gap (results) -0.089 760 ------------------------------------------------------------------------ 2. POSITIVE CONTROL -- plain per-game means (no importance split) ------------------------------------------------------------------------ metric split-half r n season/season r n mean wpa/gm [CONTROL] +0.646 2251 +0.612 2275 mean xgwpa/gm +0.859 2251 +0.775 2275 mean resid/gm +0.133 2251 +0.099 2275 ------------------------------------------------------------------------ 3. NULL SIMULATION -- shuffle HIGH/LOW labels within player-season, 200 reps ------------------------------------------------------------------------ metric test obs r null mean null 95% env p(|null|>=|obs|) clutch_gap split-half +0.047 -0.001 [-0.041,+0.044] 0.035 clutch_gap season/season -0.028 +0.000 [-0.057,+0.060] 0.358 xgwpa gap split-half +0.079 +0.001 [-0.052,+0.050] 0.010 xgwpa gap season/season +0.054 +0.003 [-0.061,+0.061] 0.119 wpa gap split-half +0.055 -0.000 [-0.045,+0.040] 0.015 wpa gap season/season +0.031 +0.001 [-0.059,+0.057] 0.259 ------------------------------------------------------------------------ 4. VERDICT ------------------------------------------------------------------------ Positive control: plain per-game wpa persists strongly (split-half r=+0.646, season-over-season r=+0.612) -- the machinery has ample power to detect a real skill of this shape. Treatment: clutch_gap season-over-season r=-0.028 (n=1749) is squarely inside the shuffled null [-0.057,+0.060] (p=0.36). The odd/even split-half r=+0.047 nominally edges the null [-0.041,+0.044] (p=0.035), but odd/even halves share calendar structure -- importance rises late in the season, so any within-season drift in team context (trades, goalie changes, form) correlates the halves mechanically, and the label-shuffle null destroys exactly that timing link. The early/late robustness split, which removes the shared timing, gives r=-0.109 (n=760). With 6 metric x test combinations examined, one p=0.035 is unremarkable. VERDICT: clutch is not a detectable persistent player skill in this data. Whatever a player's results-above-chances edge in high-importance games was, it carries essentially zero information about next season's edge (r ~ -0.03 vs a control r of +0.61), and the small within-season correlation is attributable to shared season context rather than the player. The same holds for the process (xgwpa) and results (wpa) gaps. ------------------------------------------------------------------------ 5. CAREER CLUTCH_GAP LEADERBOARD (descriptive; min 150 games, min 30 in each bucket) ------------------------------------------------------------------------ rk player pos GP hi_n lo_n gap pp/gm 1 K. Lankinen G 162 60 42 +11.215 2 I. Shesterkin G 263 87 90 +10.687 3 S. Skinner G 210 80 60 +9.721 4 P. Grubauer G 182 56 60 +9.562 5 A. Vasilevskiy G 292 94 94 +9.399 6 C. Atkinson R 176 89 35 +9.066 7 M. Benning D 157 46 48 +8.160 8 J. Swayman G 222 76 81 +7.986 9 R. Sandin D 329 114 101 +7.923 10 Z. Hyman L 348 141 93 +7.865 11 K. Guhle D 203 74 54 +7.732 12 C. Miller D 231 62 84 +7.593 13 A. Peeke D 338 116 118 +7.065 14 M. Joseph R 296 87 98 +6.986 15 W. Kaiser D 166 34 70 +6.983 Read as: extra realized-over-expected win-probability flow (percentage points per game while on ice) in high- vs low-importance games, 2021-26 pooled. Goalies dominate: the on-ice wpa-xgwpa residual is mostly saves above/below expected, and a goalie's biggest games are also the starts a coach hands him when he is running hot -- within-player selection, not clutch. Given the null persistence result, this list describes who HAS delivered results beyond chances in big games (variance's honor roll), not who WILL.