Worth waiting for, hardly worth waiting: data latency and the timing of regional landslide warnings in China

Abstract. Landslide warning services must weigh the better evidence that later data bring against the lead time lost while waiting for them. We studied this trade-off for 280 rainfall-induced landslides in China (2015–2024) with minute-, hour- or period-level occurrence times reported in the news, using the Global Precipitation Measurement (GPM) mission's Integrated Multi-satellitE Retrievals for GPM (IMERG) precipitation and as-issued Global Forecast System (GFS) forecasts in a latency-aware retrospective replay. In a case-crossover design, observed rainfall replacing forecasts added substantial information (matched area under the receiver operating characteristic curve +0.058 [0.034, 0.081]), whereas later satellite runs and newer forecast cycles added little or none. In a rolling replay of 24 h warnings, waiting one 6 h step before issuing gave no stable gain in detection at warning loads of 0.5–10 % of cell-time. In a cost–loss analysis, issuing at once with satellite rainfall plus forecast reached a relative economic value (REV, dimensionless) of 0.55, close to a latency-free benchmark (0.57) and above a rainfall threshold (0.34) and, on matched 2015–2020 decisions, the archived version 2.0.0 Landslide Hazard Assessment for Situational Awareness (LHASA) nowcast (0.34). A value-of-information waiting rule added +0.002 [−0.005, +0.007] over the best fixed policy, averaged across the exploratory cost–loss range. The pre-specified cost–loss test could not discriminate, because almost no rule warns at its cost–loss ratios, but a second test, fixed before an independent 2008–2014 sample was assembled, detected a value of waiting when no forecast was available (+0.03 [0.01, 0.05] in REV). Brackets give 95 % confidence intervals. With both satellite rainfall and forecasts available, waiting rarely repaid its lead-time cost.