Conventional wisdom, graded

As a description, it holds and it is not close. A series' goal margin splits exactly into territory and two goaltending terms, and the goaltending terms carry 77.3% of it. The club that won the goaltending won 80 of 105 series; the club that won territory won 68. When the two disagreed, 52 times, the goaltender took 32.

As a plan, it fails. The club whose starter had been the better goaltender that regular season won 49 of 95 series, 0.516; on a career-to-date figure, 41 of 80. Home ice on the same series won 0.579. The widest third of the gaps between two starters went 15 of 31 — the biggest edges did worst, which is what a coin looks like. You cannot pick the goaltender in March, and the reason is measured below.

The modern version of the goaltending story: what beats you in April is not a good goaltender but an unfamiliar one, unreadable for a fortnight and ordinary ever after. It would be worth a great deal if it were true, because unfamiliarity is knowable in March and goaltending is not. It is not true, on four measurements that would each have shown it. A club that has already seen a man this season does no better against him the next time: −0.022 ± 0.073 goals per 100 shots per meeting already played, on 18,723 goaltender-games, absorbing who he is and when in the season it was. Inside a series, where the same shooters see him four to seven times in a fortnight, he does not fall off either: across 115 men who played both halves of a series that reached six games, the front half against the back is −0.57 ± 0.43 — and the sign is the wrong way round, the back half being the better one.

Nor is a strange man any harder to solve. In his first ten NHL games a goaltender saves −0.10 ± 0.21 per 100 shots against his own season's league, and in his next twenty −0.48: the unknown man is not better, he is slightly worse, which is what you would expect of someone playing because somebody is hurt. And in April, of the two starters in each of 105 series, the club whose goaltender the other side had seen less of won 48 of 105; by NHL games played, 53; a man under a hundred games against an established one, 13 of 27. What cannot be predicted about a goaltender is not a book being written. It is the thing itself.

They were brutalised on the scoreboard and the scoreboard is not their fault. Save percentage fell in every season of this window, 0.914 in 2017-18 to 0.897 in 2025-26, 17.2 points per thousand. Over the same seven years the EXPECTED save percentage — what this book's own shot table says the shots they faced were worth — fell 19.5 points, further than the men did. Shots on goal per game fell 3.1, from 28.0 to 24.9: fewer shots, worth more each. The league did not stop stopping pucks. It started taking better ones.

And the goaltender's grip on a series has not loosened with it. The 77.3% share above, taken season by season, sits inside its own bootstrap interval in 6 of the seven; the most recent postseason reads 70.7% [44, 93], with the club that won the goaltending taking 10 of 15 series. One season of fifteen cannot tell you the weather has changed, and this one does not even try.

They are a shorter one. The same three shares, from the same rows, at three lengths: territory takes 14.8% of one regular-season night, 22.7% of a series, and 55.6% of an 82-game season, with goaltending taking the rest (85.2%, 77.3%, 44.4%). A series sits almost where a single game sits. What decides April is what decides a Tuesday in November, weighted by how few games there are.

The club with home ice won 63 of 105 series, 60%. "Always the home club" is a rule you can apply in your head, and at the series level it calls more series right (60%) than the expected-goals rating does (57.1%); the rating earns its keep only on the confidence it attaches, log score 0.664 against 0.673. In the regular season the home club wins 53.7% of 6,519 games, and the model prices it: at home in April it said 55.4% and the home club took 55.2%.

Less often than in a game, not more. The favourite in a playoff game was priced at 58.1% and won 54.6% of 306, 0.9 standard errors below its price. The club with home ice — the higher seed, which is who "the better team" usually means — was priced at 53.9% for the series and took 60.0%, a gap of 6.1 points that is 1.36 standard errors and would need 16 postseasons to settle. The regular-season standings predict a series at +0.342 log-odds per standard deviation of points spread — real, and weaker than home ice.

A game mostly is. Over 6,519 regular-season games the model is calibrated — it says 54.3% for the home club and the home club wins 53.7% — and it recovers 2.5% of the uncertainty in who wins. The rest is the puck. A series is less of one: the rating scores 0.664 against a coin's 0.693, with an AUC of 0.62 where a coin is 0.5. Knowable, thinly.

Best-of-seven arithmetic on unchanged ratings gives the Game 1 winner 66.2% of the series. Over 105 series he took 63.8%, 0.5 standard errors below the arithmetic. Winning the opener is worth exactly one game, and slightly less.

Six is the commonest length and not the majority: from the deadline, 35.7% of 56 first-round series went six, 30.4% went seven, 19.6% five and 14.3% four. Where it holds is the pool: writing "six" beside every first-round pick collected 6.0 of 24 length points an April and the model's own best answer collected 5.1, 0.6 standard errors apart. The model over-says five (26.5% said, 19.6% happened); "six" is as good as knowing.

5.97 goals a game across 105 playoff series, against 6.01 in the regular seasons of 2017-18 to 2022-23 and 6.12 in 2023-24 to 2025-26. A tenth of a goal. The model's own rating says a fifth fewer goals than are scored, in both regimes — the goals run 1.19 times its number in the regular season and 1.15 times in April — and the win probability is right anyway, which is why the scale sits on the goal draw and not in the model.

It depends on the pool. With 100 in it, whoever held the winning score had the Cup winner in 94% of 20,000 tournaments an April, and 93% of the entry's chance of first place rides on the Final. With 10 in it the winner had him 50% of the time and the entry finishes first 15.7% of the time without him. The Cup is 100 of the 380 points on winners either way; what changes is how many holders there are to beat.

Nobody's favourite has won. The crowd's champion pick — the club most brackets in the league's Bracket Challenge chose — won 0 of 4 Aprils; the model's, 0 of 4. Over the 7 Aprils the pool has priced, the club at the top of the standings won 0, the model's chalk 0, the entry 0. The crowd's share holding the eventual champion ran from 3.3% to 10.3%.

The crowd reads the standings, not the model, and it reads them well. On the 32 first-round series ESPN's crowd priced, its log loss was 0.568 against the model's 0.593 and the raw table's 0.569; a coin is 0.693. Each called 24 of 32 favourites right. The difference is 0.2 standard errors, so the honest reading is that the crowd is as good as the model, and the entry's edge is positional — where it stands in the room — not informational.

Only where the room is crowded. Against 99 table-reading entries the bracket that maximises first place finishes first 6.0% of the time, the model's chalk 1.4%, the table's own bracket 1.0% (2025-26). Against 9 the three are 23.6%, 19.6% and 16.8%. And the contrarian edge is almost entirely one pick: 93% of it rides on a champion the entry prices at 13.8%.

On 6,519 regular-season games the expected-goals rating scores 0.678 against 0.688 for goals for and against; on 105 playoff series, 0.664 against 0.679. Better in both regimes, by a little. It is not that expected goals are goals — goals run 1.19 times what the rating says — it is that the shortfall is the same for both clubs and cancels.

Measured as repeatability, because a trait repeats. 575 consecutive-postseason pairs from 294 men, each April's lift against his own regular-season rate: the correlation between one April and the next is 0.060, interval -0.029 to 0.147. 67 cleared their own rate in both Aprils against the 66 that independence predicts. On points per game instead of per sixty it is r = 0.001 over 251 pairs; of the 80 who beat their rate once, 25 did it again (31.3%), against 35.1% for everybody.

What is real is the fall: the average man scores 0.334 points per sixty fewer in April, and a skater's postseason rate runs 0.87 times his regular-season rate. The men we can all name are the ones who stayed level while the rest fell. Corey Perry's +2.91, the largest observed lift in 5 postseasons, shrinks to an estimated -0.23. Nobody elevates. Some decline less.

A starter one standard deviation above the league is worth about 12 goals a season, on a split-half reliability of 0.246. Year to year the same man's figure carries at r = 0.089 ± 0.081 on 160 pairs; his regular season into his own postseason at r = 0.063 ± 0.107, against a ceiling of 0.235 set by how few shots a postseason contains. The ceiling is the news: a perfect instrument could not read a goaltender in April at better than 0.23, and the one available reads him at 0.06.

Last season's line is a good projection and a weighted history is a slightly better one. Rank correlation with the season that followed, skaters, fantasy points: 2024-25 0.748 against 0.706; 2025-26 0.742 against 0.727. For goaltenders the naive line is as good or better: 2024-25 0.317 against 0.290; 2025-26 0.251 against 0.303. The fantasy board says so beside every goaltender.

Earlier, on this measurement. The year-over-year change in a skater's scoring rate, fitted by age on 2,207 pairs of consecutive seasons, crosses zero at 25: the climb is over by the mid-twenties and what follows is a plateau that tilts. By 30 the fitted loss is 7% a year, by 33 9%, by 35 10%. Measured on men who played both seasons, which is the men who were kept — the curve for everybody would tilt sooner.

They do better at every age, and mostly by climbing more when young; whether they decline more slowly is not shown. Same pairs, skating measured on the earlier season by NHL Edge top speed (league percentile, 2,203 pairs). Under 25 the fastest third gains +8.0% a year to the slowest third's +1.0%. Over 30 the fastest third's year is −6.0% to the slowest third's −10.4% — a gap of 4.5 points with a standard error of 3.6, 1.26σ, which is not a finding. The interaction of speed with age is -0.0014 ± 0.0002 (-5.6σ): the fast man's edge narrows with age rather than widening, though it has not closed by 33 (fitted −8.7% against −10.6%). By the report's skating grade, a judgment, there is nothing: over 30 the gap is 2.5 ± 3.6. Top speed is partly ice time, and ice time is partly quality, so some of "fast" here is "good"; and the pairs are men who played both seasons, which flatters the old in every third alike.

What the fast do is last. Of men with thirty games at 30 or older, the fastest third played thirty games the next season 96.2% of the time (151 of 157) to the slowest third's 90.1% (272 of 302), a gap of 6.1 points ± 2.3; from 33, 49 of 49 against 107 of 118. Not from a higher place: at 30 and over the fastest third was scoring 0.460 a game to the slowest third's 0.457. The old fast man is kept at the same production, which is a fact about what the league keeps and not about the curve.

It depends entirely on which thing. Year-over-year correlation of the measured traits, men clearing the floor in consecutive editions: workhorse 0.78, disciplined 0.75, slot-man 0.69, sheltered 0.66, hard-minutes 0.66, perimeter 0.66, drives-play 0.39, durable 0.32, hard-shot 0.21, clutch 0.06. Ice time, discipline and where a man shoots from are properties of the man. Whether his far shots go in (farConversion, 0.21) and when he scores (lateShare, 0.06) are mostly the season he happened to have.

It tells you who winds up. Across 5 seasons a skater's registered maximum shot speed correlates with how many slap shots he took at r = 0.55 to 0.63, and with every other shot he took at 0.13 to 0.18. A maximum records the one shot type that produces maxima. Two men with the same wrist shot sit forty feet apart on that list if one of them takes the point on the power play.

Five roster metrics, their directions written down before any was scored, on 105 series leave-one-season-out. The bar was to beat "always the home club" on log score in all 7 postseasons; the best managed 5 of 7. The strongest, age exposure, came in at z = -1.87 — and the registered null, homegrown share, a metric chosen because it means nothing, came in at z = -1.79 beside it. Five tests expect one false positive. The house prior, that a blue line that scores wins, was tested like the rest: z = -0.64, which is not that defencemen hurt but that the measure says nothing either way.

From the dead-puck floor, yes: 6.12 goals a game in 2023-24 to 2025-26 against 5.32 in 1997-98 to 2003-04, and 6.16 in 2025-26. Against the game your father watched, no: 7.84 in 1980-81 to 1985-86. The record's high is 9.58 in 1919-20 and its low 2.92 in 1928-29.

Shots on goal fell from 62.5 a game in 2017-18 to 2022-23 to 57.6 in 2023-24 to 2025-26, while scoring held (6.01 to 6.12) and the shooting percentage rose from 9.6% to 10.6%. The break is inside the tracked window, so every shot-volume figure on this site straddles it. Why is not known, and this page does not guess.

Penalty minutes a game, both clubs: 48.5 in 1986-87 to 1992-93, the roughest era in the modern window, against 17.7 in 2023-24 to 2025-26. About a third of what it was.

What is measured is the crowd: the room pays a club's top scorer above its place in the standings, and that term fits the crowd's picks. Whether the club with the star wins more series than its standing says is a different question and has not been asked of outcomes.

The feed this site reads does not say who took a draw, so no report here claims it and no probe tests it.

Blue-line scoring share said nothing about a series either way (above). Size, as such, has not been put to a series outcome, and the sample — a few dozen small defencemen playing meaningful April minutes in seven postseasons — could only find a large effect.

The game model prices a game from its two clubs and nothing that happened in the previous ten minutes. Nothing on this site has measured whether it should.