The forecast error variances used in the Goddard Earth Observing System(GEOS) global data assimilation system are currently estimated on-line by updating the time-mean statistics of observed-minus-forecast residuals. This ensures that the prescribed variances are consistent with the actual data residuals, but only on time scales longer than a week or so. To represent error variability on synoptic scales, we are developing a flow-dependent prognostic forecast error variance model. This model describes the evolution of potential temperature error variances. It involves advection by the forecast model winds and parameterized growth due to model dynamics and physical forcing. In this presentation we describe results of comparing this model with the ensemble spread implied by NCEP ensemble forecast products. We will also present preliminary results of analysis produced by the analysis system of the NASA/Data Assimilation Office with this new forecast error variance.