Operating a water supply system requires a variety of decisions to be made for different lead times and under different uncertainty conditions related to the randomness of rain events and the partial understanding of the hydrological processes. Anticipating reservoir inflows weeks to months ahead is crucial to prevent water shortages, increase systems efficiency and protect freshwater sources. In this study, we investigate the use of the Extended Streamflow Prediction method (ESP) together with the GR2M hydrological model to support the operation of the Serra Azul reservoir, a drinking water supply reservoir located in Southeast Brazil. The hydrological model was calibrated using a manual split-sample calibration–validation procedure covering the period 1979–2016 adopting the Inverse Function (F). The analysis of model performance during calibration and validation was performed using the Nash–Sutcliffe efficiency criterion (NS) and flow duration curve analysis. The NS criterion obtained during the calibration period was 0.87 and during the validation period was 0.94, results that agree with other studies in the literature. Inflow forecasts were generated for the entire period and their quality was evaluated using conventional statistical metrics and reliability and ROC diagrams. The reliability diagrams showed curves slightly steeper than the 1:1 line, indicating over-confidence for low forecast probabilities and small confidence for higher forecast probabilities. In all cases, ROC diagrams showed good discrimination between true and false alarms; the results were satisfactory comparing to the literature. The influence of the ENSO on the inflow forecasts and its impact on the reservoir operating policies were also investigated. Reservoir operation scenarios were simulated considering different initial conditions and withdrawn volumes as well as forecast inflows with different exceedance probabilities (25, 50 and 75%). The ENSO phases did not result in significant gains for the discharge forecast. Although differences were found in the values of the discharge forecasted with and without resampling, such variations had little impact on the operation of the reservoir. Finally, this study shows that there is information gain for decision making in using ESP and climate information to anticipate monthly reservoir inflows in the study area. It demonstrates that long term probabilistic forecasts have great potential to support decision making in the operation of water supply reservoirs.