Skip to content

Climate

The climate-risk primitives the IDE's climate tooling uses: reanalysis ensembles, empirical and Gumbel return periods, parametric trigger design, and average annual loss.

Reanalysis ensembles

cl = sc.sample("climate")            # Pretoria, Jan 2024 — ERA5 / MERRA-2 / JRA-3Q
sc.ensemble(cl, "t2m")               # per-date mean, spread, min, max
sc.anomaly(series, by="month")       # anomaly vs a monthly climatology

cl <- sc_sample("climate")
sc_ensemble(cl, "t2m")
sc_anomaly(series, by = "month")

          date       mean    spread   min   max
0   2024-01-01  23.400000  0.300000  23.1  23.7
1   2024-01-02  24.700000  0.360555  24.3  25.0
...
— Ensemble · t2m · 3 members
  · Median member CV 0.016: the reanalyses agree closely on this variable.

Member columns are found by name (era5, merra2, jra3q, ncep, cfsr, nora), optionally filtered by a variable prefix; pass members=[...] when yours are named differently. The note grades the agreement — agree closely below 5 % median member CV, disagree materially above 20 % — because an ensemble that disagrees is a finding, not a nuisance.

Return periods and triggers

sc.return_period(annual_losses)              # empirical + Gumbel, 2y … 500y
sc.parametric_trigger(losses, p=0.9)         # {'trigger': …, 'cap': …, 'attachment_probability': 0.1}
sc.aal(event_losses, frequencies=freqs)      # or aal(history, years=40)

sc_return_period(annual_losses)
sc_parametric_trigger(losses, p = 0.9)
sc_aal(event_losses, frequencies = freqs)

   return_period  empirical     gumbel
0              2  33.510821  33.693711
2             10  48.777226  50.338781
3             25  64.774252  58.716448
4             50        NaN  64.931480
— Return periods
  basis: 40 values over 40 years
  · Empirical: Weibull plotting position (T+1)/rank with log-linear interpolation,
    blank beyond the record; Gumbel: method-of-moments fit to annual maxima.

The empirical column goes honestly blank beyond the length of the record — a 40-year history cannot witness a 1-in-100 — while the Gumbel fit extrapolates, and is only computed when the losses really are annual maxima. parametric_trigger sets the attachment at the chosen exceedance probability with a cap at a multiple of it: the IDE's parametric design, as one function.

Function list

ensemble return_period parametric_trigger aal anomaly — each with the sc_ twin in R.