This study aims to improve stream-ow forecast at Reynolds Mountain East watersheds, which is located at the southernmost of all watersheds in Reynolds Creek Experimental Watershed Idaho, USA. Two separate models, one for the annual data and the other for the seasonal (April-June) data from 1983-1995 are tested for their predictability. Due to the difculties in collecting data during winter months, in particular the snow water equivalent (SWE), this study evaluates the impact of excluding this variable. Our results show that multilayer perceptrons (MLP) and support vector machines (SVM) are more suitable for modeling the data. The results also reveal that the difference between stream-ow forecast via annual and seasonal models is insignicant and for longer term predictions SWE is a strong driver in the stream-ow forecast. Principal Component Analysis (PCA) and Particle Swarm Optimization (PSO) are also used in this study to identify useful features. The results from PCA derived models show that PCA helps reduce prediction error and the results are more stable than using models without PCA. PSO also improved results; however, the set of selected attributes by PSO is less believable than given by PCA. The best prediction is achieved when MLP model is implemented with attributes generated by PCA.