Everything else gave way
This is not a site that shrugs. Over 6,519 out-of-sample regular-season games the model on the game pages is calibrated — what it says happens at the rate it says — and it calls 57.2% of them against 53.7% for the rule always take the home club. It is also thin: on the Murphy decomposition it recovers 2.5% of the uncertainty in a hockey game, and the honest way to say that is that hockey games are mostly not predictable and the number knows it.
April has been measured the same way, and it has given up most of its folklore. Home ice won 63 of 105 series and beats the rating at calling them. The higher seed took 60% of series where the rating said 53.9% — a 6.1-point miss, which sounds like a finding until you ask for its error bar and get 1.36σ. Five pre-registered roster shapes — age, concentration, blue-line share, buried youth, homegrown — were fitted against 105 series and every one came back null, the largest of them at 1.87σ. Nobody elevates in the playoffs: across 575 man-postseasons a player's April against his own regular season correlates at r = 0.06 [-0.03, 0.15], and the average man scores -0.33 of a point a game less than he did in January. Everybody declines. Some decline less, and we call those men clutch.
So the record is not one of a project failing to find things. It is a project that has found what it went looking for nearly every time, and been able to put a number on how sure it is.
And then there is the goaltender
A playoff series' goal margin splits exactly, with no residual, into three terms: territory — the gap in the chances each club generated — and the two goaltending terms, what my man saved above what the shots he faced were worth, and what theirs failed to. Over 105 series the goaltending terms carry 77.3% of the margin and territory carries 22.7%. The club that won the goaltending won 80 of 105 series; the club that won territory won 68. Where the two disagreed, 52 times, the goaltender took 32.
The size of the thing is the point. Series to series the goaltending term swings with a spread of 5.23 goals and the territorial term 3.06. A seven-game series is decided by a handful of goals, and one of these two terms is nearly twice the size of the other. The folklore has this exactly right, and it is not close.
It cannot be read in March
Now the other half, which the folklore does not have right at all. A goaltender's own regular season predicts his own postseason at r = 0.063 ± 0.107, on 111 starts. Year to year, his own next regular season: r = 0.089 ± 0.081. Split his season in half by alternate appearances and the halves agree at 0.246 — that is the instrument's own reliability, measured on both sides, and no correlation anywhere in this project can beat it.
Applied, it does what those numbers promise. The club whose starter had been the better goaltender that regular season won 49 of 95 series; home ice on the same 95 won 55. On a career figure instead of one season, 41 of 80. And the widest third of the gaps between two starters — the series where you were surest — went 15 of 31, the worst of the three bands. The biggest edges did worst. That is what a coin looks like when you sort it.
The ceiling is the finding
It would be comfortable to call that a measurement problem — better tracking, a better save model, more seasons, and the goaltender comes into focus like everything else did. He does not, and the reason is arithmetic rather than equipment. A postseason contains a few hundred shots. Even a perfect reading of a goaltender's true ability could correlate with what he actually does in one April at no better than 0.23, because most of what happens in a postseason is the sampling, not the man. The instrument available reads him at 0.06. The gap between 0.06 and 0.23 is what better data could buy. The gap between 0.23 and 1 is closed to everybody, for ever, and it is most of the distance.
This is the part worth sitting with. The blind spot is not where the flashlight has not been pointed. It is a place where there is nothing to see in advance, no matter who is looking or what they are paid.
Four ways out, all closed
If the goaltender cannot be picked, the natural next move is to find the thing standing behind him that can be. Four candidates have been put through a probe, each of which would have rescued the folklore as a plan rather than a description, and each of which failed.
He rises to the occasion. Nobody does, goaltenders included:
r = 0.06 on 575 man-postseasons, and the average player declines.
(npm run probe:clutch)
The right roster protects him. Five pre-registered shapes,
105 series, all null; home ice beat every one of them.
(npm run probe:signatures)
Nobody has a book on him yet. The most promising of the four,
because unfamiliarity is knowable in March. 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. He does not
fall off inside a series, where the book is written in front of us:
−0.57 ± 0.43 from the front half to the back, across 115 men who played
both halves of a series that went six, and the sign is the wrong way round. In his first ten NHL games a
goaltender saves −0.10 ± 0.21 per 100. And in April, the club whose goaltender the
other side had seen less of won 48 of 105.
(npm run probe:book)
Then he must at least be getting worse. He is not. Save percentage fell 17.2 points per thousand across this window, 0.914 to 0.897, which is the thing everyone noticed. The expected save percentage — what this book's own shot table says the pucks he faced were worth — fell 19.5 points beside it, further than he did, while shots on goal per game fell from 28.0 to 24.9. Fewer shots, worth more each. The league did not stop stopping pucks; it started taking better ones. And his grip on a series has not loosened with it: the 77.3% share sits inside its own bootstrap interval in 6 of the seven seasons taken one at a time.
What this is not
It is not that the goaltender does not matter. Goaltenders differ: a starter one standard deviation above the league is worth about 12 goals over a season, which is a real player and a real contract. The claim is narrower and worse. He matters enormously, he differs meaningfully, and the difference is small relative to what one April does to him — so the thing that decides your series is real, and reading it is beyond the instrument and always will be.
Nor is it a licence to stop measuring. Every number on this page exists because something was measured, including the four failures, and the failures are the most useful part: without them the honest sentence would be we have not looked, which is a different and much weaker thing to say than we looked four times, and here is the size of what was not there.
What you do with a blind spot
You price around it, and you wait for it to fill in. Those are the two moves, and both are built into the pool page.
Pricing around it means paying for what can be seen — home ice, the table, the room's habits — and refusing to pay for goaltending you cannot read, on either side. A bracket that leans on a hot goaltender is leaning on a number with an error bar wider than the number.
Waiting means using the one piece of evidence that is worth more than everything known in March: the games themselves. A single first-round opener moves a club's series odds by 13.9 points on average, and a bracket filled after the openers calls 68 of 105 series against 60 before them. Most pool deadlines fall after six to eight of the eight first-round openers have been played. That window is worth more than any goaltending read you could construct, precisely because the thing you cannot forecast you can sometimes simply watch.
Why it is worth playing
There is a version of this site that finds everything, and it is a duller one. If the goaltender were readable, April would be an arithmetic problem with a known answer and the pool would be a formality — the room would converge on it and the prize would go to whoever had the best spreadsheet, or a coin among the people who all had the same one.
Instead the most important factor is genuinely unknown to everyone. That is not a failure of the measurement; it is the measurement's clearest result, and the one it has tested hardest to break. The rest of the work — the reports, the ratings, the room — is about being right on everything that can be known, so that when the unknowable thing lands where it lands, you were not also wrong about the rest.
The claims behind this page are graded one by one on conventional wisdom; the measurements and the scripts that wrote them are on the record.