Probabilistic predictions for meteorological droughts based on multi-initial conditions

•A novel dynamical-statistical approach for probabilistic drought forecasting is presented.•We developed a global probabilistic drought seasonal predictions system.•Dynamical and statistical forecasts are combined with an ensemble of observations to predict meteorological drought.•Initial condition uncertainty is explicitly characterized through an ensemble of observed data.•The seasonal predictions show sufficient skill up to four months ahead for many applications. Seasonal forecasts of meteorological drought can aid decision-making in various sectors but must be trustful and skillful. One of the major drawbacks of such forecasts lies in the inherent uncertainty associated with near-real time monitoring of precipitation. This study explores the predictability of the standardized precipitation index (SPI) on a global scale combining 11 datasets as observed initial conditions with empirical and dynamical precipitation forecasts. Empirical predictions are derived from resampled historical data, while dynamical predictions rely on ECMWF’s new generation seasonal forecast model. As anticipated, the skill of SPI predictions varies depending on the target season, location, and the assessed lead times. In nearly all geographical regions and throughout all seasons, a statistically significant level of predictive skill is observed when assessing lead times spanning 2 months, with a global median correlation from 0.79 to 0.91 depending on the target season. As expected, the 4-months prediction performed worse, with a global median correlation from 0.51 to 0.77. Also, the skill is typically greater in the winter hemisphere compared to the summer hemisphere, indicating that both systems show better results in forecasting the less rainy periods of the year. The dynamical forecasts show higher performance over tropical regions and in identifying drought occurrence, especially at a 4-months lead-time. These findings suggest that SPI prediction skill is primarily influenced by the initial conditions. Better forecasts are achieved by using the complete ensemble of diverse monitoring datasets as initial condition, rather than merging the forecast with the individual products. Since all the data are available in near-real time, our results provide a basis for the development of a global probabilistic drought seasonal forecast product.