Accurate daily river flow forecasting in humid catchments with dynamic hydrological regimes is essential for sustainable water management, flood mitigation, and hydrological planning. Although machine learning (ML) and deep learning (DL) models are increasingly used for their ability to model complex, nonlinear hydrological processes, systematic comparisons of their performance—especially at daily resolution in humid basins—remain limited. This study addresses this gap by evaluating Support Vector Regression (SVR), Convolutional Neural Network (CNN), and Generative Adversarial Network (GAN) using 50 years (1969–2018) of daily meteorological and streamflow data from the Tajan River basin in northern Iran. Input variables included precipitation, evaporation, and lagged discharges (Qₜ₋₁ to Qₜ₋₃), forming five scenarios based on Pearson correlation. After comprehensive preprocessing and a 70:30 train–test split, models were assessed using R², RMSE, PBIAS, and KGE, with results visualized through scatter plots, time series, violin plots, and Taylor diagrams. SVR in Scenario 5 (SN5)—incorporating all lagged flows—achieved the best performance (R² = 0.850, RMSE = 5.675 m³/s, PBIAS = 0.475%, KGE = 0.877), significantly outperforming CNN and GAN. Notably, the DL models failed in simpler scenarios lacking lagged discharge (R² < 0.04), underscoring that input structure outweighs algorithmic complexity. The findings affirm that, under data-scarce and climatically stressed conditions, model selection must prioritize hydrological relevance over algorithmic novelty—providing critical guidance for developing reliable, operational forecasting systems and resilient water governance in vulnerable regions.