Montevideo seasonal floor¶

This record selects the temporal control that a later graph experiment must beat.

Decision¶

At tinygrad revision 6ea7d366, a train-only node mean for each ordered hour-of-week phase is the strongest validation baseline. It reduces validation MAE by 19.0% and RMSE by 27.1% relative to persistence.

Only that selected baseline is evaluated on test. It reduces test MAE by 22.4% and RMSE by 32.8% relative to the persistence result recorded by the preceding forecast study.

This is evidence for a strong periodic node field, not for graph structure. The source has ordered hourly rows but no timestamp field. t mod 168 is a periodic hypothesis; it is not a claim about a civil weekday or local time.

Protocol¶

The one-lag signal pairs the observation at raw hour r with the target at raw hour r + 1:

signal row r       x[r] ----------------------> y[r]
raw hour             r                           r + 1

target rows        0:520       520:594       594:743
                   train       validation     test

Five controls are fixed before validation:

zero            0
persistence     x[r]
daily           x[r + 1 - 24]
weekly          x[r + 1 - 168]
hour_of_week    mean(y[train rows with (r + 1) mod 168 = phase], axis=time)

The phase mean falls back to the node's overall training mean when training has not observed that phase. All 168 phases occur in the pinned training rows, so the real run uses no fallback.

Validation RMSE selects one control. MAE and then the displayed baseline order break exact ties. Test is read only for the selected control.

Result¶

CPU and Metal return the same values. Every validation candidate covers all 74 rows, 49,950 node targets, and a zero fraction of 0.824324.

The selected hour-of-week control covers all 149 test rows and 100,575 node targets:

The weekly naive control already beats persistence on validation. Averaging each node's repeated training phases removes additional noise without an optimizer.

Leakage boundary¶

Daily and weekly lookups use only raw hours before their target. Validation and test may read preceding observations across a split boundary; no later target enters a training statistic.

The phase table reads only y[0:520]. A perturbation test changes every test row while leaving the fitted phase table, all validation metrics, and baseline selection unchanged. A separate periodic fixture matches an independent Python calculation.

Reference boundary¶

The comparison is revision-bound:

PyG Temporal defaults to four lagged features and standardizes the complete series in its loader. Tinymesh retains raw one-lag values so split ownership, train-only fitting, and passenger-count metrics remain explicit. This topology-free control needs no PyTorch, NumPy, geo adapter, or new Tinymesh primitive.

Consequence¶

The previous recurrent models were compared against a floor that was too weak. The follow-up delayed-edge study subtracts this prediction and tests real directed edges against reverse-edge and permuted-field controls. It fails the promotion gate, so Tinymesh adds no edge-field API or model.

Limits¶

This result covers one pinned 31-day ordered series, one-step targets, one 168-hour hypothesis, and per-node phase means. It does not establish calendar semantics, seasonality outside this interval, moving buses, route travel time, topology value, geometry value, or production accuracy.

Reproduce¶

DEV=CPU uv run --locked python -m unittest tests.test_montevideo_seasonal
DEV=METAL uv run --locked python -m unittest tests.test_montevideo_seasonal
uv run --locked python -m experiments.run montevideo_seasonal DEV=CPU
uv run --locked python -m experiments.run montevideo_seasonal DEV=METAL