2026-27 Team Projections

Measured, not asserted: every figure below comes from backtesting on 10 held-out season transitions, 298 team-seasons back to 2010-11.Reliability is the share of one season's spread between clubs that is real rather than sampling noise. It is derived, not fitted: a kill rate is a binomial over about 250 opportunities and a faceoff percentage a binomial over about 4,800 draws, so the noise term is known and can simply be subtracted. Dividing the year-over-year correlation by it recovers how persistent the ability is, separately from how well one season measures it — which is why the penalty kill reads 0.33 and 0.99: almost perfectly repeatable, and almost invisible in a single season. That is the reason it averages three.
* marks a column whose number comes from the schedule simulation. The error quoted there is what a simple regression achieves, not the simulation's — measuring that would need a past season for which we hold both the schedule and the projections made beforehand, and no such season exists in the data. Its engine has been measured, though, which is a narrower claim and a real one: feeding the ratings 25-26 actually produced into the real simulation on the real schedule reproduces that season's points at a correlation of 0.944, a bias of under a point, and an error of ±4.43 against 9.37 for a Pythagorean on the same inputs — so the schedule is doing real work. The simulated spread is 0.873 of the real one, which looks like an engine that bunches clubs together and is not: a simulation's mean is expected points and a season is expected plus luck, and adding back the engine's own luck term (8.28 points) accounts for the difference. The compression is the luck.

Record, points and goal rates are a game-by-game simulation of the real schedule. The power play, penalty kill and faceoff dot are regressed by how much each actually carries over year to year (0.33, 0.25, 0.64 measured across 93 team-seasons); the power play also uses the season before last, which predicts it better.
The small figure beside each is the change from last season, coloured by whether it is an improvement — which for goals against and nothing else means going down.
The changes marked ~ are not forecasts of movement. The power play is a single-season regression, so its change is exactly (league average − last season) × (1 − r) — a fixed function of how far the club started from average. Sort by either and the changes come out in perfect order, on all 32 rows, because they are a rescaling of the column you sorted. A large positive number there means “they were well below average”, not “we expect them to improve”. They are greyed rather than removed because the size of the correction is still worth seeing.
The penalty kill averages three seasons, so its change depends on more than last season alone and it can reorder the league — measured on 298 team-season pairs back to 2010-11, three seasons beat one by 4.3% out of sample, because a single season measures a penalty kill so badly that only a third of the spread between clubs is real. Faceoffs are not regressed at all any more — they are built from the centres actually projected to take the draws, each shrunk by how many he took, so that column can reorder the league. Goals for, goals against and points come from the schedule simulation and are roster-driven, so those changes are real predictions too.
Sort on any column; which one matters depends on the format you play, so none is promoted over the others.