Model error is a key factor in forecast uncertainty. In a realistic case, it is unlikely that model error can be represented exactly by a physically based scheme. An alternative approach is to treat model error as unknowable and use data assimilation techniques to deduce information about the model error from observations. This poster describes this alternative approach to evaluate the effect of model error by using an ensemble of data assimilations to represent realisations of a stochastic model which contains a stochastic term defined by model errors.We can create a data-driven representation of error growth by setting the stochastic forcing term to a random draw from an archive of analysis increments with stationary statistics. Results show that the randomness assumption of the analysis increments is accurate to within 20% although the model resolution is low, approximately 125km.