The assimilation of radar observations for numerical weather prediction has drawn a good deal of attention in recent years. Although some progress has been made, there still remain great challenges. In this study, a storm-scale numerical model is used for initialization and prediction of a hail-producing supercell storm observed during STEPS. The numerical model is initialized by the WSR-88D radar data located at Goodland, Kansas using a 4D-Var technique. The results from the assimilation and prediction experiments are compared with a multiple-Doppler synthesis and reflectivity observations. It is shown that the 4D-Var technique is able to provide dynamically consistent initial conditions that lead to a 2-hour forecast with reasonable accuracy. We are currently working on the estimation of radar observational errors and their impact to data assimilation and subsequent forecast. Results will be presented in the conference.