RX5day annual maximum (1995-2014 drainage-capacity reference)

Source: ERA5-Land 0.1° (Copernicus Climate Change Service / ECMWF); horizon bands by IPCC Atlas CMIP6 delta Year: 1995-2014 Category: Physical risks · Direct Physical Loss engine inputs Coverage: Global (land only) Format: Raster grid (~9 km) Used in risk analysis: Yes — gates Mitigating services (Extreme precipitation) at ≥ 182.1 mm

What it shows

How much rain falls, in millimetres, during the wettest five consecutive days of a normal year — averaged over the twenty years of the reference period. That quantity is a standard climate index, the annual maximum 5-day precipitation, abbreviated RX5day.

How it is built

The baseline band is observed: ERA5-Land total precipitation on its 0.1° (~9 km) global land grid, summed over a 5-day rolling window labelled by its last day, taken at its annual maximum, and averaged over the twenty years 1995-2014. That average of twenty annual maxima — the mean annual maximum — is the same quantity the IPCC Atlas publishes, not a single twenty-year maximum.

The six horizon bands (2035 · 2050 · 2080, two scenarios) are delta bands. The IPCC Atlas ratio between a horizon and its own baseline is formed at the Atlas resolution, where it is a smooth dimensionless field, then resampled and applied to this observed baseline. What gets interpolated is a ratio, never millimetres.

Inside the loss engine the baseline band plays a second role: the local drainage design storm is derived from it, so a site whose network matches its own climate reads no loss however wet that climate is. Because the horizon is the baseline times a ratio, the baseline cancels exactly in the loss — which is also why the seven bands are only ever deployed together.

How to read it

Read it as a level in millimetres. Two cautions.

Resolution. At ~9 km the spatial pattern is observed rather than modelled — relief, coastlines and foehn effects are visible where a coarse model band flattened them. Three limits travel with it anyway: ERA5-Land is land only, its days are cut on UTC (a shower straddling midnight is split — negligible over five days, not over one), and 9 km still smooths convection. It names a landscape, not a rooftop, and it is not a site design storm.

It is a hazard index, not a loss. A lot of rain is not a lot of damage: drainage is dimensioned to the local climate, so a monsoon site and a temperate one are not comparable on absolute millimetres alone. The loss model prices the ratio of the storm to the local design storm, never the absolute figure.

Class thresholds

A site is flagged on Mitigating services when its wettest five days reach 182.1 mm — the point at which the loss engine's own rainfall trigger reaches half of its scale.

How this threshold was calibrated

182.1 mm is not a round number chosen for the flag. The engine converts rainfall into a triggering intensity between 0 and 1 — the same one the landslide trigger reads — and a unit test pins that intensity to exactly one half at 182.1 mm. The flag reuses the model's own midpoint. It covers 4.09 % of land pixels, cos-weighted.

Re-anchored from 170 mm on 2026-09-09. 170 mm was the same midpoint read on the previous, coarser model band; the calibration decision behind it was a share of land — a flag no louder than the cyclone or heat flags. The observed raster measures the same quantity but resolves it differently: the median moves by ×0.97, the p99 by ×1.25 (199.8 → 249.1 mm), because a 9 km grid keeps peaks that a 115 km grid smoothed away. Holding 170 mm would therefore have widened the flag from 4.09 % of land to 5.37 % — not because it rains more, but because the map sees more sharply. 182.1 mm is the measured iso-population threshold: it carries the decision already taken across the change of raster, and what actually moves is which cells are flagged.

The midpoint has two consumers and both moved together on purpose — this flag, and the rainfall trigger inside the landslide loss. Moving one without the other would have fired the landslide trigger over 31 % more land with nobody having decided so. The steepness of the curve is unchanged: only the anchor was measured.

Reading the flag on absolute millimetres is legitimate here and only here: a screening flag states a fact about the hazard and charges nothing, whereas the loss curve has to stay relative to the local design storm — charging on absolute millimetres would read a wet climate as permanent damage.

Why this layer

Extreme precipitation is priced as its own risk in the Resilience module but had no proximity flag. It overlaps the flood and storm flags by construction — pluvial and fluvial are one storm at two scales — and is kept distinct anyway because water ingress and scour have their own mechanism and their own site-type exposure.

The France-specific RX1day layer is untouched by the re-anchoring: different index, different raster, its own midpoint.