Early movement, not magnitude, forecasts the expansion of hydrological droughts

Preprint. This manuscript has not undergone peer review. Forecasting whether a young drought will grow into a major event is central to early warning, yet operational drought prediction is based almost entirely on conditions at fixed locations rather than on the drought as an evolving feature. We use a global catalogue of drought events for 1980-2024, tracked as spatiotemporal objects through precipitation, atmospheric demand, soil moisture and runoff, to ask whether the first 14 days of an event predict its later growth. The type of information that carries this skill differs between the atmospheric and terrestrial stages of the cascade. For precipitation and atmospheric-demand droughts, later expansion is predictable from the magnitude of the event. For runoff droughts, magnitude is uninformative (AUC 0.53), and the predictive signal instead lies in the early trajectory (AUC 0.71, p < 5x10^-4), and in particular in how far the object has already moved. The gain is robust to held-out regions and decades, to a range of confounds and to forward-in-time validation, and is largest in the tropics. Regenerating the catalogue on independent data indicates that the signal is confined to hydrology with pronounced storage memory, for which the failure of magnitude to predict expansion serves as an observable marker. This deposit contains the manuscript PDF and a supplement with the full analysis code, all result files, the manuscript figures, and the drought-object catalogues regenerated from independent data using an independent implementation of the 3D-DBSCAN tracker.