Time zones

Before controlling for team strength, a visitor with an inferred clock shift of 3 hours or more west loses 61.96% of the time; one with that shift east loses 54.41%, a 7.6-point swing around a 58.39% baseline.

visitor's trip              games    home win %

all games in the era       28,876       58.39
east ≥ 3h                   1,643       54.41
east ≥ 3h, short rest         457       53.61
west ≥ 3h                   1,635       61.96
west ≥ 3h, short rest         428       66.59
no long shift either way   25,598       58.41

Before reading that as jet lag, notice that it points the wrong way. Circadian disruption is supposed to punish eastward travel hardest, and here the eastward visitors do better than everyone else while the westward ones do worse. This is a reason to examine which teams make each kind of trip before attributing the difference to jet lag.

A 3-hour westward trip is, almost by definition, an Eastern-conference team visiting the Pacific coast. A 3-hour eastward trip is the reverse. The two cells are not two treatments of a comparable population. They are different sets of teams, and the home sides differ in quality accordingly.

strength edge to the HOME side, by the visitor's trip
(positive = the home team was the better side)

east ≥ 3h        −0.0281      home win 54.41%
west ≥ 3h        +0.0218      home win 61.96%
no long shift   −0.0021      home win 58.41%

The strength edge flips sign with the direction of travel, and the win rate follows it. Team strength is therefore a confound in the raw 7.6-point swing. The held-out test below asks whether direction adds predictive information after those differences are accounted for.

Denver and Utah are in Mountain time, about two hours from Eastern, so a 3-hour threshold excludes them by construction. Overlap between these terms and the model’s visiting-altitude term is 0.0%.

The test asks whether adding these terms improves predictions for held-out games, season by season, using the same walk-forward protocol ADR 0006 used.

held-out log loss (lower is better)

strength only                    0.62381
+ the four fatigue terms         0.62116     +0.00265
+ east/west × short rest         0.62119     −0.00003

Adding all four candidates changes held-out log loss by −0.00003, a small deterioration. The sign-clamped fit pinned d_east3_short at zero in 16 of 16 folds, because the unconstrained estimate pointed toward an eastward flight on short rest being an advantage.

The westward short-rest term, d_west3_short, has a mean weight of 0.0993 in 16 of 16 folds, with a coefficient of variation of 0.41. By the stability standard ADR 0006 set, that is a stable term.

It is worth −0.00002 when added to the baseline on its own. The reason is in the overlap:

d_west3_short     428 games
                    86.4% are ALSO back-to-backs
                    0.0% are also visiting altitude

Most of these games are already covered by the back-to-back term. A stable coefficient does not establish an independent contribution. The single-term comparison measures what the candidate adds beyond the existing baseline, as required by ADR 0006.

This test did not change any constant in fatigue.ts. It also did not validate the existing directional multipliers.

shipped today:   eastward multiplier   1.25
                 westward multiplier   0.85

this test found: no added predictive value from direction

The model treats an eastward shift as 1.47× as costly as a westward one. This measurement is not evidence that asymmetry is right. The test found no added predictive value from the directional terms. The ratified constants remain in the model, but this analysis does not verify them.

The question and evaluation rule were written down before the analysis ran. The result is published even though the tested terms did not improve predictions.

• —Direction remains tied to geography. The analysis controls for team strength but does not match otherwise identical trips: a 3-hour eastward trip usually involves a Western team.
• —It cannot see rest and direction as anything but a schedule fact. Nothing here observes a flight, a departure time, a hotel, or a minute of anyone's sleep. "Short rest" is a gap between dates on a calendar.
• —It cannot rule out smaller effects. The test found no predictive improvement across 16 held-out seasons; that does not establish that the biological effect is zero.
• —It cannot speak to the playoffs. The protocol is regular-season walk-forward, and a postseason series has travel patterns and rest gaps this population does not contain.
• —It cannot test the joint effect of altitude and these long shifts. The 3-hour threshold excludes Denver and Utah, and overlap with the altitude term is 0.0%.
• —Testing additional ideas on the same corpus increases the risk of selecting a chance result. New questions need a declared evaluation protocol.