The window offset: whole-domain & era

The calendar-window recovery rests on a robust diagnostic. This page carries the depth behind it: when the offset emerged, that it holds in every timezone band and season, station by station, and across the whole 0.5° domain.

The satellite-era reveal

Why does the −23 h window unlock so much? Because the recovery did not always exist. Splitting the record by year shows the calendar-window gain appears only at the 2014/15 GPM-constellation transition, once IMERG began resolving the diurnal cycle. Before that, the UTC-day r and the best-window r are almost identical and the optimal offset wanders; after it, they separate sharply and the offset locks to −23 h.

■ peak r (window-aligned, h = h★)   ■ r at h = 0 (UTC-day archive)   ■ recoverable by window alignment
Finding Before the 2014/15 GPM transition the UTC-day r and the best-window r sit almost on top of each other, and the optimal offset h★ wanders - the sub-daily timing signal is undetectable. After it, the window-aligned r pulls clear of the UTC-day r (the shaded gain, 0.20 → 0.57) and h★ locks to −22/−23 h. The calendar-window recovery only exists once IMERG resolves the diurnal cycle.

Season and timezone

Zoom into the GPM era (2015-2021) and the −23 h peak holds in every timezone band and every season. Panel (a) is the mean r(h) per band with the inter-quartile range across stations shaded; panel (b) tracks the peak offset across the twelve three-month running seasons.

Per-station offset & lift

Every station's own best offset, and the correlation it unlocks. Median h★ is h (IQR to ); the median lift from the UTC day to the matched window is Δr = , taking the typical station's peak r to .

Stations ordered west → east (bottom → top) by longitude. Each grey dot is the UTC-day *r*; the coloured dot is the window-matched *r*; the bar between them is the per-station lift.

Gridded whole-domain confirmation

The station result holds over the whole 0.5° domain, not just at the gauges. Re-window the half-hourly IMERG-L against the gridded CPC-UNI analysis: at CPC's native UTC labels the best offset is ≈ 0 h everywhere (yellow); harmonise CPC to the local-observation day and it jumps to ≈ −23 h everywhere (purple). Uniform and era-stable.

(a) CPC at native UTC labels - h★ ≈ 0

(b) CPC harmonised to the local-observation day - h★ ≈ −23 h

Gridded r(h): gauged vs whole domain

Pooling the offset sweep confirms the two levels: over the gauged cells the window-matched r reaches the per-station ≈ 0.57; over the whole domain it settles at ≈ 0.34, diluted by the data-sparse east. The peak offset is identical in the GPM and TRMM-input eras.

Finding Native CPC-UNI shares IMERG-L's UTC day, so its peak sits at h = 0; harmonised to the local day it moves to h = −23 - the same clock offset the gauges show. Gauged cells reach , the whole domain . Solid = GPM era, dashed = TRMM-input era: the offset is era-identical, so the calendar mismatch is a fixed convention, not a retrieval artefact.
Every curve here is computed with the same window-offset diagnostic that produced the [calendar-window dial](./window) - the recovery is real, robust across space, season and satellite era, and reproducible from the released code.