← Research

Playoff Translation Study -- what carries from the regular season to the playoffs?

Produced by playoff_translation.py (2026-07-08) from the new playoff PBP + shifts files (1,395 games, 16 postseasons 2010-11 .. 2025-26). Playoff attempts are priced by the season-matched regular-season xG v2 model (Fenwick + flags); player on-ice 5v5 requires exactly 5 skaters a side from playoff shifts; analysis domain is periods 1-4 (multi-OT excluded from rates). Companion data: playoff_player_gaps.csv.

# Playoff Translation Study -- console log
(playoff_translation.py, run 2026-07-08; 1,395 playoff games 2010-2025 playoffs; xG v2 Fenwick+flags built on the 16 regular seasons; on-ice domain periods 1-4.)

==========================================================================
1. ENVIRONMENT SHIFT -- does playoff hockey tighten, and how?
==========================================================================
Domain: periods 1-4 (regulation + first OT); 5v5 = situation code 1551;
shots priced by the season-matched regular-season xG v2 model (Fenwick+flags).

era           games  xG/60 att/60  xG/att  clean  reb%  rush%  PP/60 pen/60  GF/60
----------------------------------------------------------------------------------
2010-13  RS    4410   2.35   41.2  0.0571 0.0500  2.8%   2.2%  11.75   8.82   2.32
2010-13  PO     354   2.28   41.6  0.0550 0.0485  2.7%   2.5%  12.23   9.38   2.15

2014-17  RS    4961   2.36   41.7  0.0565 0.0502  2.7%   2.0%  10.79   7.69   2.30
2014-17  PO     351   2.24   41.5  0.0541 0.0482  2.6%   2.4%  10.49   8.19   2.13

2018-21  RS    4533   2.60   42.3  0.0614 0.0550  4.1%   2.0%  10.41   7.10   2.53
2018-21  PO     346   2.56   42.3  0.0605 0.0539  4.2%   2.5%  10.74   7.98   2.29

2022-25  RS    5248   2.59   42.3  0.0613 0.0581  5.8%   2.3%  10.17   7.30   2.55
2022-25  PO     344   2.49   42.2  0.0591 0.0558  5.5%   2.8%  10.41   8.49   2.33

ALL RS        19152   2.48   41.9  0.0591 0.0535  3.9%   2.1%  10.75   7.70   2.43
ALL PO         1395   2.39   41.9  0.0572 0.0516  3.8%   2.5%  10.98   8.51   2.22

(PP/60 = two-sided PP minutes per 60 min of play; pen/60 = penalty events per 60. Mean game length, periods 1-4: RS 60.8 min, PO 62.4 min; 5v5 share of play 78.0% -> 77.9%.)

Decomposition of the 5v5 xG/60 shift (playoff vs regular, pooled):
  xG/60 -3.4%  =  attempts/60 -0.0%  x  xG/attempt -3.3%
  xG/attempt split: clean-shot quality -3.6%; rebound share 3.9%->3.8%, rush share 2.1%->2.5%
  PP time: 10.75 -> 10.98 min per 60 (+2%); per game 10.90 -> 11.41 (playoff games run longer); penalty events 7.70 -> 8.51/60
  League finishing vs xG (all situations): regular -0.00 goals/100 attempts (calibrated ~0), playoff -0.33

Playoff environment by round (pooled 2010-2025):
round   games  xG/60  att/60  xG/att  reb%  PP/60 pen/60
round 1   742   2.39    42.4  0.0564  3.8%  11.29   8.81
round 2   376   2.42    42.0  0.0577  3.6%  10.63   8.14
round 3   182   2.39    40.5  0.0591  3.7%  10.54   7.95
round 4    95   2.31    40.1  0.0576  3.8%  10.71   8.77

==========================================================================
2. PLAYER-LEVEL TRANSLATION -- what regular-season signal carries?
==========================================================================
Sample: 696 skaters with >= 300 pooled playoff 5v5 minutes (of 1505 with any).
Outcome: pooled playoff 5v5 on-ice xG% (share of on-ice xG). Predictors: same-season
regular-season relWPA components per 60 (playoff-TOI-weighted across runs) +
shrunk finishing GAx/100. Weighted by playoff 5v5 TOI; standardized betas.

predictor                    weighted univariate r
EV offense /60                              +0.362
EV defense /60                              +0.072
PP value /60                                +0.191
PK value /60                                +0.102
finishing GAx/100 (shrunk)                  +0.107

WLS: playoff xG% ~ components + finishing   (n=696, weighted R^2=0.207)
term                         beta (xG% pts / SD)       t
intercept (mean xG%)                       50.37   413.2
EV offense /60                             +1.86    10.9
EV defense /60                             +1.10     7.9
PP value /60                               -0.07    -0.5
PK value /60                               +0.07     0.6
finishing GAx/100 (shrunk)                 -0.05    -0.4

Directional checks: r(reg EV offense, playoff 5v5 xGF/60) = +0.642;
                    r(reg EV defense, playoff 5v5 xGA/60) = -0.487  (defense component is +good, so negative = carries)

Grit check (run level, 2021-2025 runs with >= 50 playoff 5v5 min, n=1294, R^2=0.125):
  EV offense /60              beta  +2.16  t  +9.2
  EV defense /60              beta  +1.37  t  +7.3
  PP value /60                beta  +0.25  t  +1.2
  PK value /60                beta  +0.42  t  +2.3
  finishing GAx/100 (shrunk)  beta  -0.01  t  -0.1
  grit (hits+blocks)/60       beta  +0.25  t  +1.2

PP translation: r(reg PP value/60, playoff PP on-ice xGF/60) = +0.649
  (n=653 skaters with >= 20 pooled playoff PP minutes; compare EV: r=+0.642).
  Playoff PP TIME per 60 of play: 10.98 vs 10.75 min (+2%) -- the per-60 skill carries; how much PP opportunity exists is an era/officiating question, not a player one.

==========================================================================
3. IS 'PLAYOFF PERFORMER' A REAL ARCHETYPE? (persistence across runs)
==========================================================================
Runs with >= 8 playoff GP: 2278 (962 players).
Run-level regression for the expected playoff xG% (n=2267, R^2=0.104).
Consecutive-run pairs within player: 1316.

metric                              obs r     n null mean      null 95% env      p
on-ice gap (xG% - expected)        +0.114  1307    -0.005   [-0.057,+0.042]  0.005
finishing gap (PO-RS GAx/100)      -0.015   952    -0.007   [-0.108,+0.075]  0.736
raw playoff xG%  [control]         +0.180  1316    -0.004   [-0.057,+0.058]  0.005

VERDICT: raw playoff xG% persists across a player's runs (r=+0.180) -- that is ordinary talent, and the machinery has power. The on-ice GAP persists at r=+0.114 (p=0.005), nominally beating the null, but the regular-season control it is residualized on (relWPA components, run-level R^2=0.10) is an imperfect proxy for on-ice xG% talent, so the residual retains general skill, not necessarily playoff-specific skill. Where the regular and playoff measures are in IDENTICAL units (Fenwick finishing GAx/100, same model family both sides), the playoff-minus-regular gap has zero persistence (r=-0.015, p=0.74). Read: most of the 'playoff riser' signal is measurement leakage of ordinary talent; any true playoff-specific trait is at most weak (bounded above by r~+0.11, ~1% of gap variance).

Descriptive career on-ice gap leaderboard (mean run gap, >= 3 qualifying runs, >= 40 playoff GP; given section 3, read as variance's honor roll):
  Taylor Hall               L   runs  3  GP  41  gap +9.89 xG% pts
  Shawn Thornton            L   runs  3  GP  52  gap +9.35 xG% pts
  J.T. Compher              L   runs  4  GP  54  gap +8.57 xG% pts
  Artturi Lehkonen          L   runs  4  GP  59  gap +8.29 xG% pts
  Nick Holden               D   runs  3  GP  42  gap +7.44 xG% pts
  Blake Coleman             L   runs  3  GP  57  gap +7.34 xG% pts
  Ryan McLeod               C   runs  3  GP  52  gap +7.18 xG% pts
  Nazem Kadri               C   runs  3  GP  41  gap +7.08 xG% pts
  Andre Burakovsky          L   runs  7  GP  83  gap +6.98 xG% pts
  Alex Tuch                 R   runs  3  GP  56  gap +6.96 xG% pts
  ...
  Brad Richards             C   runs  4  GP  78  gap -8.42 xG% pts
  Josh Bailey               R   runs  4  GP  54  gap -9.48 xG% pts
  Tanner Glass              L   runs  3  GP  47  gap -11.29 xG% pts
  Mika Zibanejad            C   runs  4  GP  58  gap -11.83 xG% pts
  Nate Thompson             C   runs  5  GP  68  gap -12.96 xG% pts

==========================================================================
4. TEAM LEVEL -- predicting series winners, and playoff home ice
==========================================================================
Series with complete data: 240 of 240 (A = home-ice team; predictors = A-minus-B).
Home-ice team won the series 56.2% of the time.

model                              acc(sign)  logloss  LOSO-CV
home-ice only (intercept)              56.2%   0.6853   0.6895
regular-season points diff             56.2%   0.6675   0.6787
relWPA-sum strength diff               56.7%   0.6678   0.6790
points + relWPA                        61.3%   0.6566   0.6757

Coefficients: relWPA model logit = +0.186 + 0.014 x (wins/82 diff); points model logit = -0.202 + 0.0483 x (points-pace diff).

Per-game home win%: playoffs 53.1% (1395 games); excluding the 2019-20 bubble 53.9%.
Regular-season home win%: 54.1% (19,152 games, 2010-2026, incl. OT/SO).

Wrote playoff_player_gaps.csv (1047 players with >= 100 pooled playoff 5v5 minutes).