I've been testing dlm and KFAS packages in R to produce simulated versions of a time series. When the model is simple (say a trend and possibly a seasonal components), the simulations conditional on the observations seem to produce paths different (though apparently similar in structure) to the original time series. However, when I improve the model to avoid residual autocorrelation by introducing a new structure (say for instance an ARMA component, but in some cases it suffices to introduce a seasonal component), the simulations are disconcerting. The deviation of the paths with respect to the original data are almost imperceptible! Of course I can run an unconditional simulation, and the paths will clearly deviate from the original series. But that's not acceptable when the original series is nonstationary. So the question is, why in complex structural state space models the conditional simulations (as obtained by Bayesian inference by the dlm package, or by the Durbin and Koopman methodology in KFAS) produce paths that are almost exactly equal to that of the original series?