Data-driven machine learning approaches have advanced Soil Moisture (SM) forecasting at 1–14-day lead times using observational time series, but the hydrological controls on spatially heterogeneous forecast skill remain poorly understood, hindering targeted model improvement. Here, we show that spatial heterogeneity in data-driven SM predictability is associated with site-specific hydrological characteristics, particularly soil texture and precipitation extremes. Using a stacking ensemble, we established an empirical benchmark of SM predictability at 1–14-day lead times across three soil depths at 73 monitoring sites in Guangdong Province, China. Forecast skill separated into two regimes (slow-decay v.s. fast-decay), characterized by fast or slow decay as lead time increased, respectively, reflecting contrasting soil water dynamics. Event-scale diagnostics showed that fast-decay sites underwent stronger vertical redistribution and deeper SM responses during extreme precipitation, eroding the SM memory that supports longer-lead forecasts. A joint threshold of profile-wide silt content (>70 %) and extreme daily precipitation (>28 mm/d) delineated that fast-decay regime. We synthesized these characteristics into Fast-Decay Index (FDI), which summarizes the local hydrological conditions associated with fast loss of predictability. Because its inputs are available prior to model training, the FDI has the potential to support pre-modeling assessment of SM predictability at new sites, pending independent-site validation. The FDI provides a physically interpretable diagnostic for comparing site-specific predictability and may inform targeted model development and observation design.