Fit, measured
fit: not yet validated
Nothing called fit appears on a page until its pre-registered thresholds pass. Until then the components show as description with "fit: not yet validated".
P7.1 Point-in-time snapshots and leakage canary
passed its pre-registered test but is not adopted (owner decision pending) · registered 2026-10-08 · run 2026-10-08
P7.2 Move backtest
passed its pre-registered test but is not adopted (owner decision pending) · registered 2026-10-08 · run 2026-10-08
P7.2 v2 Move backtest on point-in-time RAPM
did not pass its pre-registered test · registered 2026-10-08 · run 2026-10-08
P7.4 Lineup optimiser: projected minutes
did not pass its pre-registered test · registered 2026-10-08 · run 2026-10-08
P7.5 Goalie fit
did not pass its pre-registered test · registered 2026-10-08 · run 2026-10-08
P7.2
Did the fit components predict what happened after a move?
Every regular's club change from 2010-11, against matched players who stayed, with propensity weights; flops counted as outcomes; selected on 2010-2021 and proven once on 2022-2025.
P7.2 v2
The same test, on a measure club strength cannot move
The first test's outcome was measured against the club without the player, so a stronger or weaker new club moved it on its own. The second test kept the moves, the stayers and the components, and scored each player's change in point-in-time RAPM (5v5 xG per 60, adjusted for teammates and opponents), using only games before the move for the before value.
FAIL: fit surfaces stay off; published as a finding.
P7.4
Projected minutes against the first 20 games
The lineup optimiser places each mover in his new club's best lineup on what was known the day before; the projection is that slot's minutes.
The optimiser did not pass, and it did worse than carrying a player's own minutes over, so no page shows its projected minutes.
P7.5
Goalie fit
A goalie's Gaussian-process smoothed save surface integrated against the new club's shots against, compared with what he did there.
P7.3
Head coach for every game since 2010-11
From NHL game-centre feed (right-rail gameInfo.headCoach). A game the feed does not name a coach for reads unknown.
- 2010-11: 2638 of 2638 club-games named
- 2011-12: 2632 of 2632 club-games named
- 2012-13: 1610 of 1612 club-games named, 2 unknown
- 2013-14: 2646 of 2646 club-games named
- 2014-15: 2638 of 2638 club-games named
- 2015-16: 2642 of 2642 club-games named
- 2016-17: 2634 of 2634 club-games named
- 2017-18: 2710 of 2710 club-games named
- 2018-19: 2716 of 2716 club-games named
- 2019-20: 2424 of 2424 club-games named
- 2020-21: 1904 of 1904 club-games named
- 2021-22: 2802 of 2802 club-games named
- 2022-23: 2800 of 2800 club-games named
- 2023-24: 2800 of 2800 club-games named
- 2024-25: 2796 of 2796 club-games named
- 2025-26: 2788 of 2788 club-games named
- 2026-27: 110 of 110 club-games named
Assistant coaches are unknown: no allowed per-game source.
Interim status is not in the feed; a short tenure is shown as a tenure, not labelled interim.
Name spellings are kept as the feed gives them; variants that normalise to the same letters are merged.
Model cards
Fit study cards
fit study
Point-in-time snapshots and leakage canary (P7.1)
Guarantee that no fit input uses data from after a move.
Training data
What it learned from
Nothing to show.
Evaluation
How it did on seasons it had not seen
- sampled
- identical
- positiveControlFired
Calibration
Do 60% calls win 60% of the time?
Not on file.
Known limits
Where it stops
fit study
Move backtest (P7.2)
Decide whether anything on the site may be called fit.
Training data
What it learned from
Nothing to show.
Evaluation
How it did on seasons it had not seen
- n
- r2Baseline
- r2Fit
- deltaR2
- permutationP
- topFifthMeanY
- bottomFifthMeanY
- topMinusBottom
- band95
- T1
- T2
- pass
- flopRate
- matchedResidualMean
Calibration
Do 60% calls win 60% of the time?
Not on file.
Known limits
Where it stops
fit study
Move backtest v2: point-in-time RAPM (P7.2 v2)
Decide whether anything on the site may be called fit, with an outcome club strength cannot drive.
Training data
What it learned from
Nothing to show.
Evaluation
How it did on seasons it had not seen
- n
- r2Baseline
- r2Fit
- deltaR2
- permutationP
- topMinusBottom
- T2bar
- band95
- expansion
- T1
- T2
- T3
- T4
- pass
- flopRate
Calibration
Do 60% calls win 60% of the time?
Not on file.
Known limits
Where it stops
fit study
Lineup optimiser: projected minutes (P7.4)
Project a mover's minutes and slot at a new club.
Training data
What it learned from
Nothing to show.
Evaluation
How it did on seasons it had not seen
- n
- within2
- mae
- pass
- byGroup
- injuryBandCoverage80
Calibration
Do 60% calls win 60% of the time?
Not on file.
Known limits
Where it stops
fit study
Goalie fit: smoothed save surface (P7.5)
Predict a moving goalie's goals saved above expected per shot at the new club.
Training data
What it learned from
Nothing to show.
Evaluation
How it did on seasons it had not seen
- n
- r
- band95
- pass
- rawPriorR
- dietOnlyChange
- power
Calibration
Do 60% calls win 60% of the time?
Not on file.
Known limits
Where it stops
What fit will mean, and what it never includes
Performance fit is stated in on-ice units (isolated 5v5 xG share, points per 60, ice time) with bands. The feasibility gate (cap path, roster rules, clauses, the two front offices) is shown beside it as yes, no or unknown, and is never folded into the number. Nationality is never an input; attention (White Noise) is context only. Clusters: defencemen under 24, rookie forwards, players 33 and older, new arrivals, first 20 games.