Insights
Where the numbers say something a scoring table does not: postseason production, finishing and goaltending against expectation, and scoring by strength state.
The short version
Does the ordering hold up?
Projection ordering persists deep into the player pool
Across the five NHL seasons studied, every Points, Goals and Shots-on-Goal evaluation in ranks 101–200 showed ordering evidence. Explore how ordering changes by season, statistic and rank region.
Negative results in shallower regions are often inconclusive because those regions lack enough information to resolve the preregistered effect size — not because the ordering failed there.
Explore the ordering study →Playoff scoring, 2025-26
Points in the postseason — the games every other board on this site leaves out. Ranked on the total with games played on the row rather than on points per game: a per-game sort with no minimum puts a two-point sweep above a twenty-nine-point run, and the exposure varies more than fivefold across a field decided by elimination.
How these boards are built — regular season only, and not comparable across seasons
Against expectation
Finishing — above expected, 2025-26
Skaters who scored the most goals above what the model expected from their attempts. Our own tests find this partly carries from one season to the next, which is what makes it a reading of skill rather than of luck — though only partly, and the panel below says how much.
Finishing — below expected, 2025-26
The other tail of the same column, shown because a board that only ranks one direction implies the other end is not a measurement. It is the same statistic with the sort reversed.
Goaltending — above expected, 2025-26
Goalies who conceded fewer goals than the model expected from the attempts they faced. Attempts here are open-play and unblocked — deliberately not the box score's shots against, which counts a different set of pucks. Save percentage is not shown for the same reason: the game log already gives it over that other set.
Goaltending — below expected, 2025-26
The same column sorted the other way. A goalie on a heavily-outshot team can sit here on volume alone, which is why the attempts faced are on the row.
Has this actually worked?
The projection above says how much of a hot finishing season should carry into the next one. That is a forecast, so it can be wrong, and this is where it gets graded. Two different things are shown and they are not interchangeable.
Tested on 2025-26 — a season the rule never saw
a backtest, not a track recordThe unscoreable ones are players who did not play 2025-26. They are counted rather than dropped — a rule that only looks good on the players who came back would otherwise look the same as one that works.
2.7% lower prediction error
A small margin is the honest outcome here, not a weak one — the rule keeps under a fifth of a hot season, so its forecast is close to this baseline by construction. A large number would be evidence of a bug.
18.3% lower prediction error
The live record
Empty, and that is accurate rather than a gap. Every projection is written down the night it is made, but none has come due yet — these are forecasts about next season, so the first ones can be graded once it has been played.
The difference from the backtest above matters. A backtest can be re-run until it flatters; a prediction recorded before the season it describes cannot be, because the data it will be scored against does not exist yet.
Power-play points, 2025-26
Points recorded while the skater's side had the man advantage, taken from the on-ice strength state of every event rather than from a summary. Shots here are shots on goal — a different count from the attempts on the finishing boards above.
No club is shown, and the reason is worth stating: we hold each player's CURRENT club, not the one they played for in the season a row describes. Among the top 50 finishers that would be wrong for 8 of 50 in 2025-26 and 29 of 50 in 2021-22, so it is left off rather than shown wrong. Follow a name to their page for club and history. Every number here was computed overnight and is only read at page load, never recalculated.