Abstract
<title>Abstract</title> <p>Regulated Mediterranean basins require monthly forecasts that are useful for near-term planning and transparent about their limits. Building on a validated explanatory framework for the Jucar River Basin, this study advances from retrospective reconstruction to multi-horizon forecasting under scenario-conditioned climate and demand pathways for 2024–2030. A data-driven pipeline generates monthly forecasts for aquifer piezometric levels, reservoir storage, and river discharge by separating endogenous memory (ENDO; lagged states and seasonality) from exogenous forcing (EXOG; bias-adjusted climate predictors and sectoral demand drivers). Models are trained on monthly increments and deployed recursively across lead times up to 48 months under a leakage-aware pseudo-future protocol, with performance assessed using Kling–Gupta Efficiency (KGE) and its correlation, variability, and bias components. Hindcast results reveal clear typology-dependent skill patterns rooted in the intrinsic predictability of each system component. Aquifer forecasts maintain median KGE above 0.97 across all horizons, but their high scores must be interpreted alongside a demanding persistence benchmark because groundwater levels are strongly autocorrelated. Reservoir models substantially outperform a naïve persistence benchmark from six months onward (RMSE-based skill score up to +0.50 within the evaluated long-range bucket), while river models add meaningful skill within the 12-month reliability horizon (skill score +0.38 at six months). KGE decomposition identifies systematic overestimation, captured by the bias ratio component, as the dominant degradation mechanism for rivers at longer leads, which exogenous forcing partially mitigates. Forward experiments under three scenario families (Base, Favorable, Unfavorable) show that scenario divergence is strongest for rivers within the reliability horizon, while remaining muted for aquifers and storage-conditioned for reservoirs. Scenario spread is interpreted as sensitivity to forcing assumptions, not as calibrated predictive uncertainty.</p>