Historical hydrometric context and explainable multi-horizon evaluation of a GloFAS-derived discharge product near...

Open global hydrological products are often the only continuous source available for data-scarce locations, but their local use is stronger when the modelled signal is connected to documented hydrometric evidence. This study evaluates an explainable machine-learning workflow for anticipating the GloFAS v4-derived discharge series returned by the Open-Meteo Flood API near Babahoyo, Ecuador, and complements the modern 2017-2026 analysis with a historical hydrometric-context protocol based on official INAMHI yearbooks for the San Pablo-Babahoyo system. Direct models were assessed at t+1, t+3, t+7, and t+14 using leakage-aware features, rolling-origin backtesting, a 365-day holdout, prediction intervals, product-state classification, permutation importance, and SHAP. ElasticNetCV achieved R²/NSE of 0.900 at t+1 in the final holdout, while mean backtest skill weakened at longer horizons. The dominant role of current and recent discharge values shows that the main short-term contribution is efficient local post-processing of GloFAS persistence; meteorological and seasonal information becomes relatively more important as lead time increases. The study therefore contributes a transparent multi-horizon evaluation, an auditable pathway for historical gauge-level corroboration, and practical diagnostics for deciding when local emulation is informative and when the native global product should remain the primary reference.