Leave one city out
A complete city is excluded from fitting and uncertainty calibration.
GEOGRAPHIC ROBUSTNESS LABORATORY
Can the model travel?
This laboratory holds out complete cities and regions, measures domain shift, compares forecasts statistically and publishes worst-domain performance—not only average scores.
Leave one region out
Every city in one region becomes an untouched external test domain.
Shift diagnostics
Target shift, feature shift and event-rate shift accompany every fold.
Block bootstrap
Paired skill uncertainty is estimated without destroying short temporal dependence.
Forecast comparison
Horizon-aware Diebold–Mariano statistics compare candidates with persistence.
Robust leaderboard
Ranking includes mean error, worst-domain error, event skill and interval calibration.
What is reported
City and region holdouts
Target, features, events
Bootstrap and DM tests
Season and intensity
Research boundary. Synthetic demonstrations validate software behavior only. Real generalization claims require frozen official datasets, harmonized variables and independent replication.