The NOAA Physical Sciences Laboratory produces the Global Ensemble Forecasting System (GEFS) which comprises 11 ensemble members (1 control and 10 perturbation runs) for over a 36-year period (December 1984 to present), with forecasts initialized every day for the next 16 days (first 8-day forecasts obtained from a high-resolution grid and the next 8-day forecasts from a low-resolution grid). The system provides 36 variables related to a wide range of hydrometeorological processes. In this study, we assess the predictability of precipitation within the context of statistical downscaling using a minimum set of predictor variables (precipitation and temperature). We use feedforward backpropagation neural networks with a suite of training algorithms to determine which variables (features) are of most relevance at different forecast lead times. The outcome of this study will significantly benefit short-term flood forecasting using GEFS data.