Reservoir storage in hyper-arid basins responds to drought with a delay that is rarely quantified with observations. This study combines a monthly satellite record of the Chah-Nimeh 3 reservoir in Sistan, Iran (December 2018 – November 2025, 83 months) with local precipitation, potential evapotranspiration and the three-month Standardized Precipitation (SPI-3) and Standardized Precipitation Evapotranspiration (SPEI-3) indices. A Standardized Water-Level Index (SWLI) derived from the satellite storage is used to describe reservoir drought. Storage declined significantly over the record (Mann–Kendall Z = -5.03), from an annual mean of 108 million cubic metres (MCM) in 2019 to 47 MCM in 2023, and 13 months had SWLI ≤ −1, almost all of them in the 2023–2024 drought. Lagged correlation showed that SPI-3 leads the SWLI by 4 months (r = 0.38) and SPEI-3 by 5 months (r = 0.28), which reflects the delayed propagation of meteorological drought through the transboundary Hirmand system. Six models (MLR, SVR, XGBoost, an SVR tuned by the sparrow search algorithm, an XGBoost tuned by the whale optimization algorithm and a GRU network) were compared with persistence for forecasting the SWLI one and three months ahead under a chronological split. At one month SVR-SSA was the best model (RMSE = 0.437, NSE = 0.75) and improved on persistence (RMSE = 0.525) by 17%. At three months all models lost skill, but XGB-WOA still reduced the error of persistence by 10%. Antecedent SWLI dominated the one-month forecasts, whereas the six-month precipitation total and SPI-3 lagged by four to five months became the most important predictors at three months. The results quantify drought propagation in a data-scarce transboundary basin and provide a simple early-warning tool.