Short-Term Estimation of Standardized Precipitation Index Using Machine Learning: The Case of Çanakkale and Biga

This study attributes the apparent short-term predictability of the Standardized Precipitation Index (SPI) to machine-learning calibration, target-window overlap, climatology, and drought memory. All SPI parameters, climatologies, z-scores, scaling, and model selection were fitted within leakage-free expanding-window folds, followed by an embargoed independent test. One-month SPI-12 produced R² = 0.918 and 0.885 at the two stations, but precipitation-window climatology matched or exceeded the learned model in six of the ten tasks with R² ≥ 0.40. An effective unknown fraction was reported and six- and twelve-month sensitivity experiments reduced overlap: mean R² was 0.466 for the 50%-unknown SPI-12 (t+6) task, but fell to 0.009 for fully unknown SPI-6 (t+6) and -0.012 for fully unknown SPI-12 (t+12). Event-based validation detected only 64 of 205 observed drought events across the primary tasks, and training-only F1-based algorithm selection yielded only a small macro-recall increase. Grouped Shapley analysis assigned an average 72.8% of absolute contribution to the full-window reference block combining observed precipitation and climatology. Thus, high long-timescale scores primarily support conditional drought-continuation outlooks, not independent prediction of future precipitation anomalies or sudden drought onset.