100: League PDO (should be 100). 0.28: How much finishing repeats. 298: Shots for finishing to be half-skill. 0.98: xG calibration (actual/pred) Every offseason brings a wave of "due to regress" takes, and for the PWHL they are usually a glance at PDO (a team's on-ice shooting percentage plus its save percentage, which sits near 100 for everyone over time) and a shrug. The instinct is right. The execution is usually vibes. We wanted to do it properly: measure how much each number actually repeats in PWHL data, then shrink every result toward the mean by that amount, so the "candidates" are calibrated and not just whoever sits at the top of a noisy list. One check has to come first, because everything here leans on it. All of this compares goals to expected goals, so expected goals had better be honest. Across every shot this season, our model predicted almost exactly the goals that happened: actual finishing came in at 0.98x predicted, and the curve below tracks the perfect-calibration line bin for bin. Good. Now we can trust the gaps. Teams: who outran their chances Start with the cleaner story. League PDO is 100, as it always is, and the spread around it is mostly luck that comes back. But not all of PDO is equal. The shooting half (on-ice shooting percentage) regresses almost completely: teams do not have a repeatable knack for an extra goal per chance. The save half is different, because some of it is goaltending, which is a real and somewhat repeatable skill. So when a team is running hot, the question is which half is doing it. BOS (PDO 103.6; goals 64% vs expected 52%, +11.0) : scored well over its chances, led by goaltending (SH 8.4%, SV 95.2%).. OTT (PDO 100.9; goals 51% vs expected 44%, +6.9) : scored well over its chances, led by goaltending (SH 7.0%, SV 93.9%).. MTL (PDO 101.5; goals 60% vs expected 55%, +4.9) : scored well over its chances, led by goaltending (SH 6.9%, SV 94.6%).. VAN (PDO 100.4; goals 50% vs expected 48%, +2.9) : scored well over its chances, led by a mix of shooting and goaltending (SH 8.6%, SV 91.8%).. MIN (PDO 101.3; goals 53% vs expected 52%, +1.1) : scored well over its chances, led by shooting (SH 9.1%, SV 92.3%). BOS is the league's biggest over-performer, scoring on 64% of the 5-on-5 expected goals while only earning 52%. Some of that is a genuinely strong goaltender and will hold; the shooting-percentage part will not. On the other side, the unlucky teams should climb even if they do nothing differently: SEA (PDO 97.1; goals 38% vs expected 45%, -6.7) : scored under its chances, led by goaltending (SH 6.2%, SV 90.9%).. NY (PDO 98.1; goals 46% vs expected 54%, -8.1) : scored under its chances, led by a mix of shooting and goaltending (SH 6.3%, SV 91.8%).. TOR (PDO 97.6; goals 41% vs expected 52%, -10.6) : scored under its chances, led by shooting (SH 4.8%, SV 92.8%). TOR is the clearest positive-regression case: it created 52% of the expected goals in its games but cashed only 41%, on a 5-on-5 shooting percentage of 4.8% that almost has to rise. Short seasons make this worse, not better: with only about thirty games a team, the luck swings are large, so the pull back to the middle is stronger here than the NHL intuition suggests. Players: the finishing map, and how little it repeats Now the part everyone wants and the part that needs the most care. Finishing, scoring more or fewer goals than the chances were worth, feels like a skill, and a sliver of it is. But we measured how much a shooter's finishing in one set of games predicts her finishing in the rest of the season, and the answer is not much: about 0.28 on a 0-to-1 scale. For finishing to be even half skill and half luck, a shooter would need around 298 shots , and almost nobody in a PWHL season gets close. So the honest move is to shrink hard: take each shooter's gap and pull most of it back toward zero, leaving only the part the data says is real. Ran hot (expect a pullback). These shooters beat their chances by the most. The shrunk projection is not a prediction that they will be bad, only that this exact goal total is unlikely to come again on the same looks. Jessie Eldridge (F, 14 goals on 7.2 expected, 75 shots) : about 82% of that is the kind of thing that evaporates, so the projection is nearer 8 goals on the same chances.. Taylor Heise (F, 12 goals on 6.0 expected, 88 shots) : about 73% of that is the kind of thing that evaporates, so the projection is nearer 8 goals on the same chances.. Grace Zumwinkle (F, 13 goals on 8.5 expected, 82 shots) : about 81% of that is the kind of thing that evaporates, so the projection is nearer 9 goals on the same chances.. Daryl Watts (F, 10 goals on 6.3 expected, 86 shots) : about 75% of that is the kind of thing that evaporates, so the projection is nearer 7 goals on the same chances.. Rebecca Leslie (F, 13 goals on 9.3 expected, 98 shots) : about 77% of that is the kind of thing that evaporates, so the projection is nearer 10 goals on the same chances. Ran cold (expect a bounce). These shooters got to good spots and did not convert. Finishing that low rarely repeats either, so the arrow points up. Natalie Spooner (F, 3 goals on 7.7 expected, 75 shots) : unlucky more than bad, with a bounce toward 7 goals on the same chances.. Emily Clark (F, 3 goals on 7.5 expected, 77 shots) : unlucky more than bad, with a bounce toward 7 goals on the same chances.. Anne Cherkowski (F, 2 goals on 6.0 expected, 75 shots) : unlucky more than bad, with a bounce toward 5 goals on the same chances.. Alina Müller (C, 3 goals on 6.5 expected, 61 shots) : unlucky more than bad, with a bounce toward 6 goals on the same chances.. Laura Stacey (F, 6 goals on 8.5 expected, 112 shots) : unlucky more than bad, with a bounce toward 8 goals on the same chances. What “due to regress” does and doesn't mean One thing the loud takes get wrong. Regression to the mean is a statement about a group , not a promise about a person. Saying a shooter is due to regress does not mean she will be bad; a genuinely good finisher who ran hot is most likely to come down and still finish above average . Shrinkage does not erase a real talent, it right-sizes it. Read every arrow above as “less of this than last year,” not “this player is a fluke.” And remember the whole board nets to roughly zero: one shooter's good luck is another's posts, not extra goals in the league. What this read isn't Not a forecast for any one player. The projections are calibrated group expectations. Any individual can beat or miss hers; the shrinkage just says the smart baseline is closer to the line than to this season.. Finishing barely repeats at PWHL samples. The repeatability we measured (~0.28) is from splitting one season, which if anything overstates it. Treat the player map as low-confidence and the team PDO read as the firmer one.. Goaltending is the real part of PDO. The save half of a hot team's PDO can be a genuinely good goalie and will not fully regress; the shooting half almost entirely will. We flag which is which but do not separate a goalie's skill from her team's defense here.. Our xG, and only shot-based. Calibration is good this season, but the values are our model's, and this is finishing on shots, not a complete account of offense. Methodology Calibration. Every shot binned into ten expected-goal deciles; predicted vs actual conversion per bin. Season-wide actual / predicted = 0.98.. Team PDO. 5-on-5 on-ice shooting percentage plus save percentage (excludes empty-net and special teams by construction), shown next to the goals-share vs expected-goals-share gap.. Finishing. Goals minus expected goals, all strengths, with empty-net and penalty shots removed. The talent-vs-luck split is estimated from the league-wide spread of finishing against the binomial luck each shooter's shot volume implies (empirical Bayes); each shooter's gap is then shrunk by that, and the regressed share (60+ shot minimum) is reported as the projection.. Repeatability. Finishing per shot on odd-numbered games vs even-numbered games, correlated across shooters and Spearman-Brown corrected (~0.28). Split-half within a season is an upper bound on the true season-to-season number.. Prior art. PDO traces to Vic Ferrari's blog and Gabe Desjardins' Behind the Net; the shooting-talent-is-mostly-luck result to Tom Tango, Eric Tulsky, and Michael Schuckers; xG over-performance to Evolving-Hockey and MoneyPuck. The PWHL-first piece is measuring the repeatability in this league rather than borrowing the NHL's. Full team and player tables live under /stats ; the xG model's calibration has its own page at /stats/calibration .