Accurate discharge forecasting is important for effective river basin management, flood control, and water resource planning. In this study, an XGBoost-based multi-output regression approach is developed and evaluated to forecast river discharge up to six hours ahead using historical discharge data and engineered temporal features. The study outlet is located near Hurricane Mills, Tennessee (USGS monitoring site 03603000). Hourly discharge data from 2016–2024 were preprocessed by imputing missing values, resampling at regular intervals, and dividing into training (2016–2020) and validation (2020–2024) datasets. Feature engineering included lagged discharges, hourly differences, 24-hour differences, and time features such as hour of day. The XGBoost model was trained in a multi-output stepwise manner, with each horizon being predicted by a separate regressor and assessed according to a set of statistical metrics. Results show that the model achieved excellent predictive performance, with mean absolute error (MAE) increasing gradually from 0.99 m³/s for predictions 1 hour ahead to 4.79 m³/s for predictions 6 hours ahead. The root mean square error (RMSE) was in the same range between 4.44 m³/s and 21.06 m³/s, while the mean absolute percentage error (MAPE) was below 3% for all lead times, reflecting the model strength in capturing both short- and medium-term discharge dynamics. Also, efficiency statistics of Kling-Gupta Efficiency (KGE > 0.99 for the first three steps), Nash-Sutcliffe Efficiency (NSE > 0.99 for 1 hour ahead and 0.98–0.99 for subsequent steps), and R² values close to unity point towards very good correspondence between predicted and observed flows. Analysis of feature importance revealed that recent discharge lags and short-term discharge differences were the most predictive features, which reflected the importance of recent flow history in prediction. Overall, the proposed XGBoost-based framework provides outstanding accuracy for short-term discharge forecasting and provides an efficient data-driven approach to hydrological decision support systems in river basins.