Assessing the limits of climate-driven and memory-enhanced streamflow forecasting in the regulated euphrates river under CMIP6 scenarios

Previous Euphrates and regulated-river forecasting studies have often compared algorithms without quantifying how far prediction can proceed when reservoir releases, storage, withdrawals, and allocation rules are unavailable. This study tested that information limit for monthly discharge at Haditha, Abbasiya, Kufa, and Nasiriyah using station-specific discharge records within 1980–2023 and observed rainfall/temperature from Iraqi meteorological sources. ARIMA, ARIMAX, SARIMAX, Random Forest, and XGBoost were arranged as a stepwise information hierarchy: persistence, climate, seasonality, nonlinear learning, and open-loop recursive prediction. Five CMIP6 GCMs (ACCESS-ESM1-5, CNRM-CM6-1, HadGEM3-GC31-LL, MPI-ESM1-2-LR, and MRI-ESM2-0) under SSP2–4.5 and SSP5–8.5 were downscaled with LARS-WG 8 and used only to generate climate-conditioned scenarios for 2021–2060. Performance limits were quantified using NSE, RMSE, KGE, benchmark comparison, recursive degradation, and SHAP attribution. Statistical skill was weak except at Abbasiya (SARIMAX NSE = 0.281; RMSE = 35.98 m 3 s⁻ 1 ), while Random Forest improved one-step Abbasiya skill (NSE = 0.639) but recursive skill declined at most stations. Hydrological memory contributed 70.35%, 69.38%, 65.45%, and 43.51% at Abbasiya, Haditha, Kufa, and Nasiriyah, respectively, indicating that antecedent flow carries more predictive information than local climate within the available dataset. Projected mean changes were − 14.55/ − 14.85%, + 8.62/ + 8.40%, + 23.04/ + 23.48%, and − 17.27/ − 17.33% under SSP2–4.5/SSP5–8.5, but these are low-confidence conditional scenarios where validation was weak. For Iraqi water allocation, drought preparedness, and transboundary management, operational implementation requires release, storage, inflow, diversion, and allocation data.