Previous simulation tests have shown that MM5 has quantifiable skill in forecasting the occurrence of snowstorms in the northeastern US which typically have strong synoptic forcing. However, the timing, duration, and amount of snowfall predicted by MM5 are much less accurate. Recent efforts have shown that data assimilation techniques can be effective in improving the initial conditions and subsequent forecasts of snowfall. This work explores the feasibility of using the MM5 four-dimensional variational data assimilation (4DVAR) system to incorporate multi-radar data into a snowfall forecast. The goal of the data assimilation process is to obtain an optimal initial state that is consistent with radar observations and suitable for 1-4 hour snowfall forecasts. Observation System Simulation Experiments (OSSE) are conducted for a snowstorm event in the New York City area, using simulated data that emulate Doppler radar observations. Performance of MM5-4DVAR are evaluated, especially in terms of its ability to recover the moisture structure and reproduce the snowfall pattern and amount from Doppler radar observations. The relative importance of assimilating radar reflectivity and radial velocity are assessed. Sensitivities of the retrieval and forecasts to data coverage and observational errors are also tested.