Reliable daily streamflow forecasts are essential for flood warning and water allocation in the rain-fed catchments of central Chile, which have been affected by a decade-long megadrought. This study forecasts next-day discharge of the Cauquenes River at El Arrayán (622 km², CAMELS-CL station 7336001) for 1979–2019 using random forest (RF), light gradient boosting machine (LightGBM), long short-term memory (LSTM), a convolutional–LSTM network (CNN-LSTM) and two hybrid models in which LightGBM hyperparameters were optimized by Harris hawks optimization (LGBM-HHO) and particle swarm optimization (LGBM-PSO). Inputs were catchment-average precipitation, temperature and potential evapotranspiration from CR2MET and antecedent discharge, and the target was modelled in logarithmic space. Models were trained on 1979–2004, validated on 2005–2011 and tested on the megadrought period 2012–2019. All models outperformed persistence (test NSE = 0.69). The default LightGBM achieved the best test scores (NSE = 0.825, KGE = 0.855, RMSE = 5.22 m³/s), followed by LSTM (NSE = 0.808), whose validation performance was the best of all models. The metaheuristic searches reduced the validation loss only marginally, and the tuned models did not generalize better (LGBM-PSO NSE = 0.798; LGBM-HHO NSE = 0.755) while requiring about 20 min of computation each. The LSTM forecast high flows almost as well as LightGBM, but both deep learning models overestimated low flows during the drought. The results suggest that for daily forecasting with abundant data, well-configured gradient boosting and LSTM models are sufficient, and the added value of metaheuristic tuning is limited.