This study presents an advanced framework for streamflow prediction in the Seybouse watershed, Algeria, by applying machine learning (ML) and optimization-based hybrid models. Three modeling techniques Extreme Learning Machine (ELM), Online Sequential ELM (OSELM), and Grey Wolf Optimizer-enhanced ELM (GWO-ELM) were developed to forecast daily streamflow at two hydrometric stations: Bouchegouf and Ain Berda. Input variables were selected using a robust feature selection algorithm incorporating lagged streamflow, precipitation, and evaporation. Model performance was evaluated using R2, NSE, RMSE, and MAE metrics. Results showed that the ELM1 model achieved the best validation performance at the Bouchegouf station, with R2 = 0.567 and NSE = 0.553, while the GWO-ELM5 model performed best at the Ain Berda station, with R2 = 0.316 and NSE = 0.231. Despite high accuracy during training, the models faced performance degradation in validation, particularly under peak flow conditions. Nevertheless, all models showed limitations in capturing extreme peak flows and suffered from decreased performance during validation, especially in nonlinear conditions highlighting issues of overfitting and the need for improved generalization. These findings underscore the potential of hybrid ML approaches in hydrological forecasting while emphasizing the need for more robust strategies to improve model adaptability and accuracy in diverse hydrological contexts.