Over the past decade, ensemble forecasting has emerged as a powerful tool for numerical weather prediction. Not only does it produce the best estimates of the state, it also provides the uncertainties associated with the best estimate and the predictability of a certain event, which provide invaluable information to estimate the background error covariance for data assimilation. In this study, random perturbations have been used to initialize the mesoscale ensemble forecast of the 24-25 January 2000 “surprise” snowstorm that occurred along the East Coast of the United States. The error growth characteristics estimated from these ensemble forecasts are found to be quantitatively similar to previous studies that were achieved by perturbing the individual observations or by adding small-scale monochromatic wave disturbances. Analyses show that the isotopic random perturbations initially dissipate everywhere except for in the conditionally unstable regions. Error from the ensemble forecasts quickly develops horizontal and vertical structure over time that is associated with the rapid cyclogensis. Error growth is very nonlinear and is dictated by moist processes. These ensemble forecasts are also used to examine the flow and time-dependent background error covariance with the goal of understanding and developing an Ensemble Kalman Filter (EnKF) system for meso- and regional scale data assimilation. Preliminary results show that the initially non-correlated random errors develop strong spatial correlation not only among the same variable (auto-correlation) but also between different forecast variables (cross-correlation) within 6-12 hours especially over the region of strong cyclogenesis. When an observation is taken, these auto- and cross-correlations from the short-term ensemble forecast can spread the information for both observed and unobserved variables. The ensemble forecast can also be used to determine where the optimum observations should be taken maximizing the Kalman gain (“target observation”).