Mesoscale simulations of storm systems over the western U.S. have been conducted using the Weather and Research Forecasting (WRF) model. Composite correlation matrices have been created using the model results and corresponding re-analysis fields. The matrices indicate that there are large discrepancies between forecasts and analysis in the coastal region of southern California. These discrepancies are especially pronounced in cases of extreme precipitation over the Sierra Nevada mountains. Since the storm systems generally move in from the northwest, west, and southwest, the uncertainties and errors in the initial conditions over this coastal region lead to poor forecasts in the western U.S. A method has been developed to identify those problematical areas where the WRF model exhibits poor forecasts. The error growth of the model is calculated as a function of the forecast lead time. These statistics are used to develop a perturbation strategy, i.e., a method to create an ensemble of initial conditions that can be used to construct an ensemble forecast. To clarify the relative importance of initial conditions and physical parameterizations, ensemble forecasts are made in two modes: (1) altering initial conditions only, and (2) altering parameterizations only. The probability density functions of forecasted precipitation over the Sierra Nevada mountains for both regimes, i.e., for ensemble generated from initial conditions and parameterizations separately, are examined in an effort to clarify the relative importance of the initial analysis and parameterizations in ensemble forecasts.