Fast domain experiment

Two cheaper Ecuador configurations against the operational run — GFS-initialized 72 h hindcasts, verified against the INAMHI gauge network during an extreme El Niño.

Read the three differences differently. Fast2 − Fast1 is the controlled contrast: identical physics, vertical levels, 3 km grid and initial conditions, so the only thing that changed is whether a 15 km parent supplies the lateral boundaries or GFS does directly. The two comparisons against the operational run also change the physics (Thompson/RRTMG/YSU at 42 levels against HERN at 65) and the nest footprint at the same time, so a difference there cannot be attributed to any one cause.

Side by side

All four panels share one view and one lattice. The difference is subtracted cell by cell in the browser — the fast grid is an exact subgrid of the operational nest, so no interpolation is involved.

Press , and . to step through time; the arrow keys pan the maps.

Fast1
Fast2
Operational
Difference

Field

Difference

Skill against the gauges

Skill is 1 − MSE/MSEoperational on the matched sample. The whisker is a 95% interval from a bootstrap that resamples gauges, not gauge-hours: rain is correlated in space and time, and resampling hours would treat thousands of dependent observations as independent and make every interval far too narrow.

Where the skill comes from

Performance diagram

Bias by forecast lead

A configuration whose advantage lives only in the first hours is winning on spin-up, not on forecast quality.

Run by run

Whether a configuration wins consistently, or wins on average because of one case.

Timing of the rain

Diurnal cycle

Hourly rates by local hour, whichever accumulation the scores above are set to. A model that rains at the wrong time of day can still score well on daily totals, so the phase is worth reading on its own.

Skill against cost

The numbers

Wet bias: coverage or intensity

Which configuration wins where

Fast1 Fast2 Operational

The event these runs were scored on

Late September is climatologically the dry season on the Ecuadorian coast, but this period falls inside an extreme El Niño, and the gauges did not record a dry-season sample.

These stations' hourly records begin in September 2026, so no anomaly against a local climatology is computed here — the figures above are what the gauges measured, not a departure from normal.

Caveats

. The rain was concentrated rather than widespread, so the categorical scores at the higher thresholds rest on a small number of heavily weighted events and the intervals are correspondingly wide.

Gauges are points and the model is a 3 km cell average, so the nearest-cell match penalizes a correct storm that is displaced by one grid box twice — once as a miss and once as a false alarm. The scorer also stores 9 km and 15 km neighborhood matches; the headline numbers use the nearest cell.

Gauges excluded as broken:

Companion pages: run-by-run verification · forecast viewer