The Brazilian electricity sector is strongly dependent on hydropower and the accurate prediction of streamflow series is essential for planning and risk management. Recently, hybrid models, which combine prediction and data preprocessing techniques, have stood out. However, in the literature there is no consensus on the predictive superiority of these hybrid models versus their pure version. This paper aims to contribute to the literature with the evaluation of prediction performance suitability of pure and hybrid models for monthly stationary and non - stationary series of streamflow. For this, models were constructed using Artificial Neural Network and ARIMA forecasting techniques coupled with the Singular Spectrum Analysis (SSA) and Seasonal and Trend decomposition based on Loess (STL) data pre-processing techniques. As a result, this study shows that pure models obtained a better performance for the Belo Monte (stationary series), already hybrid models were the best for the Sobradinho (non-stationary series).