New to these reports? Start here
- Dotted-underlined words have a plain-English definition — hover or tap them. Every term is also on the glossary page.
- "Null" means we found nothing, not that something broke. Most reports here are negative results, on purpose — knowing an idea doesn't work is the point.
- Two questions get asked separately. First, is the effect real? Second, is it already priced into the betting odds? An effect can be completely real and still useless to bet on.
- A "calibration" row is a self-check. It runs the same method on something already known to be true. If that fails, the whole report is unreliable — so it's reported alongside the findings.
- If a confidence interval includes zero, the real effect might be nothing at all, so no claim gets made.
Detecting departures without an early-departure list
2026-09-08. We have no historical early-departure data, so the working premise was: read departures off game-1 rosters instead, accepting that we only learn them once the season starts and so cannot forecast with them.
Measured on college_hockey_player_game_stats (2021-2025, ~1,100 games/season)
plus the transfer sheet ingested in e5746b5. 6,637 returning-player decisions.
The premise holds, at 79%
"Absent from the team's first game" as a test for "never plays for them again":
| gone all next season | played again | |
|---|---|---|
| missed game 1 | 2,727 | 741 |
| played game 1 | 0 | 3,169 |
Recall is 1.000 by construction — never playing again implies missing game 1 — so it is not evidence of anything. Precision is the real number: 0.786. One in five players written off at game 1 comes back, and they are not marginal: median 12 games, 59% play 10 or more.
Waiting two weekends buys most of the accuracy back
Precision of "absent from the first K games":
| K | flagged | precision | false positives |
|---|---|---|---|
| 1 | 3,468 | 0.786 | 741 |
| 2 | 3,252 | 0.839 | 525 |
| 3 | 3,162 | 0.862 | 435 |
| 4 | 3,091 | 0.882 | 364 |
| 6 | 2,989 | 0.912 | 262 |
| 8 | 2,937 | 0.928 | 210 |
The true-positive count never moves (recall stays 1.0); only false positives shrink. Four games — about two weekends — halves them and lifts precision to 0.88. Eight games gets 0.93 and costs a month.
Most departures were knowable before game 1 anyway
Of 2,101 departures across 2022-2024:
| share | |
|---|---|
| in the transfer sheet | 34.3% |
| aged out (4+ seasons played) | 26.8% |
| moved but not in the sheet | 2.5% |
| left D1 with eligibility left | 36.4% |
| -> knowable before game 1 | 61.1% |
⚠ Career length is undercounted for anyone who started before our data (2020, and 2020 is only ~25% scraped). Restricting to players first seen in 2022+, whose whole career is observed, gives 56.0% knowable and 40.3% early — but that cohort cannot age out by 2024, so its 0% graduation rate is an artifact and its 40.3% is specifically the underclassman rate. The honest range is 55-61% knowable, 36-40% genuinely early.
What that means per team
Median roster that dresses: 27 players.
| per team per season | median | mean | p10 | p90 | max |
|---|---|---|---|---|---|
| departures | 11 | 11.3 | |||
| of which early | 4 | 4.1 | 1 | 8 | 13 |
Early departures are 15.2% of a roster per season.
Conclusion for the study
The pessimistic framing — "we won't know the rosters beforehand" — is worse than the data. Roughly 60% of roster churn is knowable before a puck drops, from the transfer sheet plus counting seasons played, and neither needs an early-departure list. The irreducible surprise is ~4 players per team per season, and even that resolves to 88% precision by the fourth game.
So the design should be two-stage rather than one:
- Preseason — subtract known transfers and aged-out players. This is ~60% of the churn and is available before game 1.
- From game 4 — fold in the players who never showed. Do not use game 1 alone; its 21% false-positive rate is players who go on to average 13 games.
⚠ Standing caution from the TODO, unchanged: this measures WHO left, not what their leaving did. Movement is not random and regression to the meanregression to the meanExtreme results tend to be followed by more ordinary ones, purely by chance. Mistaking this for a real decline is a classic error. will impersonate a "new team" effect for anyone who moved after a career year.