Effective lead-time stream flow forecast is one of the key aspects of successful water resources management in arid regions. In this research, we present new data-driven models based on Statistical Learning Theory that were used to forecast flows at two time scales: seasonal flow volumes and hourly stream flows. The models, known as Support Vector Machines, are learning systems that use a hypothesis space of linear functions in a Kernel induced higher dimensional feature space, and are trained with a learning algorithm from optimization theory. They are based on a principle that aims at minimizing the generalized model error (risk), rather than just the mean square error over a training set. Due to Mercer's condition on the kernels the corresponding optimization problems are convex and hence have no local minima. Empirical results from these models showed a promising performance in solving site-specific, real-time water resources management problems. Stream flow was forecasted using local-climatological data and requiring far less input than physical models. In addition, seasonal flow volume predictions were improved by incorporating atmospheric circulation indicators. Specifically, use of the North-Pacific Sea Surface Temperature Anomalies (SSTA) improved flow volume predictions.