The future is, of course, inherently uncertain, but in most practical applications it is prediction of the future that is required. There is then no possibility of constraining uncertainty with respect to future observables, and uncertainty estimation is necessarily a forward estimation process, albeit that the models used may have been conditioned on past data and performance. However, as shown by the classic post-audit analyses of groundwater models in Konikow and Bredehoeft (1992) and Anderson and Woessner (1992), future boundary conditions may be difficult to predict even if model parameters can be constrained on past data. Thus, assumptions about the choice of boundary conditions should be justified and the uncertainty associated with that choice assessed by forward propagation, taking account of the uncertainty in model choice and parameters