Philadelphia Flyers

Both bars share one scale per row — the wider of the two league ranges — so a bar that moves is a number that moved. The grey tick is that column's league average, and the arrow is the change in league rank. Goals-against and expected-goals-against rank ascending, so low is good and green always means improved.
Goals for, goals against and points percentage come from the roster: summed player projections over the real 26-27 schedule, last season's expected goals against regressed toward the league and adjusted for projected goaltending, and a game-by-game simulation of that schedule.
The special teams are projected from how much each number actually carries over year to year, measured across 298 team-season pairs going back to 2010-11. The power play uses one prior season shrunk 0.36; the penalty kill uses three, weighted .5/.3/.2 and shrunk 0.38. Special teams barely persist, so a league-worst penalty kill is mostly bad luck and comes most of the way back, while the faceoff dot is a repeatable team skill. How far a club moves depends on how far from the mean it started: Vancouver sits near league average on both the power play and the faceoff dot and barely shifts, while Edmonton's 30.6% power play comes back to the middle and this penalty kill, worst in the league, is the biggest riser of all 32.
The penalty kill averages three seasons because one measures it so badly. Split a single season's spread into ability and sampling noise — a kill rate is a binomial over about 250 opportunities, so the noise is known rather than guessed — and only 33% of the gap between clubs is real. Divide the year-over-year correlation by that and the ability underneath comes out near 1.0: penalty killing is almost perfectly persistent and merely hard to see in one season. The cure for a noisy measurement is more of it, and three seasons beat one by 4.3% out of sample. Because averaging reorders clubs, that row's rank is a real forecast; the rows marked ⁼ are single-season regressions, which compress the values but keep them in order, so their rank is last season's by construction and only the value is a prediction.
A roster-aware faceoff model was built and rejected on the evidence. Draw counts are not stored, but they can be recovered exactly from the per-game percentages, and with real draws a club's number reconstructs from its own centres. It still does not predict better: given a club's first half to learn from and even handed the second half's draw distribution, it scored 2.24 against the plain regression's 2.26 across 32 clubs, and the best blend of the two puts almost no weight on it. A team's faceoff percentage already carries who takes its draws. Expected goals has no team projection at all, so that row shows last season alone.