That's because the initial conditions are different each model cycle so the perturbation of those initial conditions among the ens members often results in somewhat different outcomes from one run to the next. The ensemble forecast is a 'check' on the higher res deterministic for that particular run- each ens run gives an indication of uncertainty(error in initial conditions) for that particular deterministic outcome, and so (especially at longer range) they will also reflect some differences from run to run. 8-10 days out I simply monitor the mean over a series of consecutive runs, and the outcome generally stays pretty consistent, depending on the model, number of members, etc. Anecdotally the Euro/EPS seems more consistent from one cycle to the next. That's my take and I am sure its not completely correct lol.