From short to long-term horizons: Comparative assessment of Machine Learning and Deep Learning for river stage prediction with spatial transferability analysis

• LGBMR achieved superior accuracy (R²>0.98) with minimal computational demands. • GRU unexpectedly outperformed LSTM for 30-day forecasts in deltaic rivers. • Systematic hyperparameter optimization revealed horizon-specific configurations. • Models demonstrated robust spatial transferability across hydrological stations. • PCA-based ranking provided objective multi-criteria performance assessment. River stage forecasting is critical for flood management in deltaic regions experiencing intensified hydrological extremes under climate change, particularly in vulnerable Bangladesh. This study systematically evaluated six machine learning and deep learning algorithms—Support Vector Regression (SVR), Random Forest Regression (RFR), Light Gradient Boosting Machine Regression (LGBMR), Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU)—against persistence baseline for multi-horizon (1–30 days) river stage predictions in Bangladesh's Old Brahmaputra River system. Using 26-year daily hydrological data (1999–2024) from Jaganathganj Station, with comprehensive transferability testing at upstream Islampur Station, all models underwent systematic hyperparameter optimization through grid search ensuring fair algorithmic comparison. LGBMR consistently demonstrated superior performance for short-term forecasting, achieving Root Mean Square Error (RMSE) of 0.271 m (R² = 0.985) for 1-day and 0.543 m (R² = 0.939) for 7-day predictions with remarkable computational efficiency. Unexpectedly, GRU outperformed complex architectures at 30-day horizons (RMSE = 0.878 m, R² = 0.841), suggesting architectural simplicity benefits when temporal dependencies become diffuse. Computational efficiency analysis revealed LGBMR required minimal resources (0.14–0.77 s training time), while deep learning models demanded 25–132 s. Principal Component Analysis-based ranking confirmed LGBMR's dominance across short-medium horizons, achieving normalized composite scores of 0.8597–1.000 at 1–7-day predictions. Hydrometeorological augmentation incorporating rainfall and temperature data yielded 13.5–17.8 % error reduction for long-term predictions, with GRU benefiting most at 30-day horizons. Model transferability assessment demonstrated robust spatial generalization, with Nash-Sutcliffe Efficiency exceeding 0.90 for short-term and 0.84 for long-term predictions. Statistical validation through bootstrapped confidence intervals and paired significance testing ( p < 0.001) confirmed all ML/DL models substantially outperformed persistence baselines. These findings challenge conventional assumptions about model complexity, revealing gradient boosting methods provide superior accuracy-efficiency balance for operational flood forecasting, while simpler recurrent architectures unexpectedly excel at extended predictions.