Data assimilation (DA) is increasingly being employed to estimate the parameters and states of terrestrial ecosystem models from eddy covariance measurements of net carbon (C) fluxes. The length of the observation time series used varies for each study. The impact of these differences has not been quantified explicitly. Therefore, in this study, we investigate the importance of the time series length relative to observation noise and data gaps. Different length synthetic time series are used to determine the parameter and C stocks of a simple ecosystem C model. Two commonly used DA schemes are tested: the sequential Ensemble Kalman Filter (EnKF) and a batch Metropolis Markov chain Monte Carlo algorithm. Longer time series improve both the parameter and C pool estimates of the EnKF, while adversely affecting those of the Metropolis algorithm. For both DA approaches, the length of the time series has more influence on the parameter and pool estimates than the level of random noise or amount of data. In this study, the EnKF provides more robust parameter and C pool estimates than the Metropolis algorithm. Optimized parameters and states are often used as the basis for forecasting future responses. Despite having better parameter and C pool estimates, EnKF forecasts estimates have much larger uncertainties than the Metropolis algorithm forecast estimates. Finally, we suggest that the structure of simple box models, as used in this study, introduces a large degree of equifinality into DA. Neither DA scheme correctly accounts for the equifinality, but our results suggest that it is particularly problematic for the batch Metropolis algorithm.