image © phloxii/iStock/Thinkstock
The philosophy of ensemble forecasting is to capture possible outcomes rather than a single definitive answer. Thus some variability in a sequence of NWP model forecasts is to be expected. Every new forecast run is, on average, better than the previous one - but it is also different. New observations modify previous analyses of the atmospheric state, and therefore forecasts generated from these new analyses evolve and develop differently to earlier results. Usually, the differences in the forecasts are no more than moderate but occasionally, differences can be quite large. These may appear as a significant forecast jump in evolution of the atmosphere (eg developing a more mobile or amplified upper trough or a significant increase in modelled cloud cover). This has a consequent effect on forecast values. This is an unavoidable consequence of a imperfect dynamical forecast system but it's not a problem as such. To the uninformed user or customer, this can be very disconcerting and cause lack of confidence. To the informed forecaster it is an opportunity to deduce additional information and produce a better probabilistic forecast arguably of more use to the customer.