4-3: Final OT (NY). 1.81-1.78: xG (NY-MIN). 0.30: Fillier 3G total xG. -2.19: Rooney GSAx With 7:40 left in regulation New York trailed 3-1, had been outshot in two of three periods (final shot share 29 to 25), and was sitting at 1.81 xG against Minnesota's 1.78. The model put the win probability comfortably with the visitors. Then Sarah Fillier scored three goals in 17 minutes and 25 seconds of game-clock, on shots the model priced at 0.07, 0.13, and 0.09 xG. Combined: 0.30 xG of total goal expectation, three actual goals . The third was the overtime winner. The first two arrived 14 seconds apart. The natural hat trick, decomposed Three goals in a row by the same skater without anyone else scoring in between is the textbook definition of a natural hat trick. Sarah Fillier 's first came at 12:20 of the third on a 0.07 xG backhand, the kind of shot the model says scores roughly one time in fourteen. Her second came at 12:34 on a 0.13 xG wrister, fourteen seconds of game-clock later. Her third came 4:45 into overtime on a 0.09 xG wrister for the win. The xG total across all three is 0.30 . That is the model saying these three shots, on average across many PWHL seasons, would produce somewhere around one third of a goal in expectation. Producing three is the kind of single-game outlier that doesn't survive contact with a 10-game rolling sample, and isn't supposed to. It is what variance looks like when a shooter and a moment both peak at the same time. Goaltending, decomposed Maddie Rooney stopped 25 of 29 live shots worth 1.81 xG in regulation and overtime, finishing at -2.19 GSAx . That GSAx number reads worse than the underlying performance was. Three of the four goals against her were under 0.13 xG, and Fillier's three in particular are the sort of below-expectation shots that tilt a goalie's single-game GSAx hard without telling you anything about how she actually played. Callie Shanahan on the other end stopped 22 of 25 live shots worth 1.78 xG, finishing at -1.22 GSAx . The three Minnesota goals were a Heise wrister 16 seconds into the game, a Zumwinkle 5-on-4 backhand, and a Zumwinkle wrister 44 seconds into the third. None were high-danger looks; the goalie line came out close to even on a similar workload to her counterpart. Minnesota's other side of the ledger Until the Fillier sequence, Minnesota was running the game. Three of their goals came on shots under 0.16 xG and the team finished above 50% Corsi for the night. Their PP went 1-for-4 ( Grace Zumwinkle 's 5-on-4 backhand at P2 4:07), and the early scoring opened a two-goal cushion that should have been enough on the model's chance distribution. The 14-second response was the kind of sequence that converts a comfortable lead into a one-goal game, and the OT came down to who finished first. The game in cumulative xG The scoring timeline P1 0:16 · Taylor Heise · xG 0.10 · NY 0, MIN 1. P1 5:20 · Maddi Wheeler · xG 0.07 · NY 1, MIN 1. P2 4:07 · Grace Zumwinkle (PP) · xG 0.15 · NY 1, MIN 2. P3 0:44 · Grace Zumwinkle · xG 0.05 · NY 1, MIN 3. P3 12:20 · Sarah Fillier · xG 0.08 · NY 2, MIN 3. P3 12:34 · Sarah Fillier · xG 0.11 · NY 3, MIN 3. P4 4:45 · Sarah Fillier (OT) · xG 0.11 · NY 4, MIN 3 What this read isn't Not a claim Sarah Fillier is on this trajectory every night. A three-goal night on a handful of shots is a finishing tail that does not repeat on demand. The right test is the season-long shooting profile against expected; one game is one data point.. Not a claim Maddie Rooney played poorly. A -2.0 GSAx night against four goals where three were under 0.13 xG is mostly variance landing on her side of the ledger. The underlying save process was much closer to expected than the headline GSAx number suggests.. Not a claim Minnesota collapsed. They produced a shot profile that the model graded as a winning process for 52 minutes. The Fillier sequence was a hot stretch by one shooter, not a structural breakdown by Minnesota.. Not a claim the model is wrong. A 60-40 favourite losing in OT to a hat trick on 0.30 xG is well within the variance the model's probability outputs allow. Single-game outcomes can land anywhere in the tails; the season-long log-loss is what tells you whether the model is right. Methodology xG model : PWHL-calibrated logistic regression on 16,709 non-EN shots. Same fit as the methodology post . Features: distance, angle, rebound (≤3 s), penalty shot, power play, overtime, scorer-tagged quality chance, seven shot-type dummies. Empty-net shots are priced outside the main model (MoneyPuck / NST convention).. GSAx : summed xG on shots faced minus goals allowed, computed per-goalie from the PBP feed. Empty-net shots and goals are excluded from both sides of the ledger.. Numbers regenerate on every build. A PBP re-sync or xG recalibration reshapes every figure in this post on the next static export. Full play-by-play, shift overlays, and the per-player game log for this game are at /games/95 .