Retraining frequency: metric drift since the last retraining

Each CV test fold is binned by elapsed days since the fold start (the retraining point), in two-week bins. The last partial week is excluded. Points are pooled over the 8 folds; the table below shows within-fold slopes averaged over folds, which removes most fold/season differences.

Models are retrained at the start of each 13-week fold. If predictions degrade noticeably with elapsed time, retraining more often may help.

Chronos-2 and Kumo are excluded: they are zero-shot and always consume history up to the prediction time, so they are never stale with respect to the retraining schedule. Any within-fold drift they show is due to other time-varying factors, not retraining frequency. TabICL (8 estimators, full-history context) is included because its in-context training set is refreshed only once per fold.

Within-fold slopes (per 30 days)

MAPE is in percentage points; pinball-d2 and coverage are absolute changes. For MAPE, positive is worse. For pinball-d2, negative is worse. Coverage is relative to the nominal 0.90 level.

Caveat: only 8 folds; the slopes are linear approximations and residual seasonal effects remain. The final partial week is excluded because it is only 7 days and is seasonally noisy.