54%: On-ice xGF%, shift fresh. 35%: By 60+s on the shift. 44s: Median PWHL shift. +6: Extra fade for D vs F (pts) Every hockey coach knows the back half of a long shift is where goals get given up. In the PWHL it should matter more than almost anywhere, because the benches are short and the shifts run long: the median shift in our data is 44 seconds, and the 90th percentile is 75. Star skaters live out there. So we asked a simple question of every 5-on-5 shot this season: who was on the ice, how long had she been out, and which way did the expected goals tilt? Stack that up across the year and you get a fatigue curve. One honest caveat first, because it shapes everything. A shift runs long for a reason. Usually that reason is that the team is hemmed in its own end and physically cannot change. So late-shift numbers are dragged down partly by tired legs and partly by the trouble that kept the players out there in the first place. We cannot fully separate the two with this data, so we never call the raw curve pure fatigue. What we can do, and do below, is measure each skater against the league's own decay, so the part that is just hockey cancels out. The league curve, and the cliff Here is on-ice expected-goals share (xGF%, the share of the expected goals on the ice that go your team's way) by how many seconds a skater has been on her shift, at 5-on-5. A team is strongest around 15-30 seconds in (54% xGF, controlling the play), then it slides, and past 45 seconds it falls off a cliff, down to 35% once a shift passes a full minute. Defenders, who get stuck out longest, fall the hardest. Notice what drives the fall. It is not that a tired team stops generating, it is that it stops defending. In the longest band of shifts, the on-ice expected goals split roughly 297 for the team to 545 against: the shots against pile up far faster than the shots for dry out. That is the signature of getting caved in, which is exactly the being-hemmed story. Read the curve as a warning about time on the ice, not as a tidy measurement of lactic acid. Forwards hold, defenders fade The clearest robust finding is positional. Forwards and defenders are nearly even fresh (53% and 54% xGF), but on long shifts the forward group holds at 41% while the defender group sinks to 36%. A defender's long shift is a different, worse animal than a forward's, mostly because she is out long for the worst reason: her team cannot get out of its own end. We keep that in mind below by grading each skater against her own position, not against the other. Who holds up, with the honest asterisks Now the individual read, and here we slow down. We measure each skater's fade (fresh xGF% minus long xGF%) against her position's average fade, so a positive number means she sustains better than a typical forward or defender who reached the same shift age. We shrink that number toward zero for small samples, and we only trust a name if it shows up in both halves of the season. The replication is real but modest: across qualified skaters the first-half and second-half signals correlate at r = 0.14. That is enough to point at the extremes with confidence and not enough to rank everyone in between. So we name the ends and leave the middle alone. Take Kendall Coyne Schofield (F) and Maggie Connors (F). Fresh, they are nearly the same skater: 61% and 60% on-ice xGF in the first 20 seconds of a shift. Past 45 seconds they are not close. Kendall Coyne Schofield is still at 56%, while Maggie Connors sinks to 33%. Same fresh number, very different tail. That gap is invisible in a season stat line and it is exactly what a bench boss wants to know before sending someone back out tired. Holds up: Sophie Jaques (D, 54 to 52), Grace Zumwinkle (F, 50 to 55), Marie-Philip Poulin (F, 54 to 58). Fades most: Maggie Connors (F, 60 to 33), Allie Munroe (D, 59 to 29), Kali Flanagan (D, 59 to 33). The sustainers are not just good players, they are players whose on-ice results barely move as the shift drags, which is rarer than it sounds. The faders are mostly defenders whose long shifts turn into fire drills. Both groups cleared the same bar: enough long-shift shots to mean something, and the same direction in both halves of the season. What a coach can and can't take from this It is tempting to turn this into a double-shift cheat sheet. Resist that. This measures what happens inside one long shift, not how fast a player recovers between shifts, and those are different things: sustaining a 50-second shift is not the same skill as being ready to go again 40 seconds later. It is also an on-ice result, shaped by linemates and matchups, not an isolated skill. The honest takeaway is narrower and still useful: when a shift has to be stretched, the personnel on the ice is not neutral. Some skaters keep the play roughly even deep into a shift; others hand the other team a grade-A look. That is one input on shift-length risk, not a verdict on a player. What this read isn't Not pure fatigue. Long shifts are not random; they happen when a team is pinned. The raw curve blends tired legs with being trapped, and we cannot fully separate them with shot and shift data alone.. The individual signal is modest. The split-half correlation is r = 0.14. The league and position curves are robust on pooled data; the named individuals are the well-sampled extremes, not a trustworthy 1-to-N ranking. Most skaters are not distinguishable from league average.. Not a recovery or double-shift metric. It is within-shift sustain, not between-shift recovery. Do not use it to decide who can go right back out.. On-ice, not isolated. Late-shift results reflect the whole unit on the ice and the opponent, not one player in a vacuum.. 5-on-5 only, our xG. Special teams are excluded (their shift dynamics are different), and the expected-goals values are our model's, with real error on any single shot. Methodology On-ice reconstruction. For every 5-on-5 shot we find the skaters whose shift overlaps the event time and compute each one's shift age as event time minus shift start, in game-clock seconds (which stop on whistles, so this is running-play time, not wall time). The shot's expected goals are credited for or against each on-ice skater at her shift age.. Bands. The league curve uses 15-second bands. The player read pools a fresh band (under 20s) against a long band (45s+), chosen to fit the PWHL's long shifts and to give individuals enough sample.. Player metric. Decay residual = (a skater's fresh xGF% minus her long xGF%) minus her position's average of the same, so forwards are judged against forwards and defenders against defenders. Positive means she sustains better than her position. The residual is shrunk toward zero by sample size.. Validation. Every skater's residual is recomputed separately in the first and second halves of the season; the two correlate at r = 0.14 across the 90 qualified skaters. Named players cleared an exposure floor (110+ long-shift shots) and pointed the same direction in both halves.. Numbers regenerate on every build. A play-by-play re-sync, a shift re-import, or an xG recalibration reshapes the curve and the named lists. This is the first of our shift-level reads. It pairs with the on-ice and line-combination tools under /stats . The natural next step is to bring in score state and zone of the shot to chip further at the being-pinned effect.