The accuracy of nowcasts based on advecting the observed rain field is limited by the extent to which the observed field develops during the forecast period. This temporal development can be either decay in existing rain bands or the development of new rain bands in areas that are currently not raining. A convenient way of representing this uncertainty in the forecast is to express the forecast as a probability distribution rather than as an expectation. This paper presents a method to generate a series of stochastic forecasts that are conditioned on the current rain field. The apparent motion of the field is first estimated as a field of advection vectors, thereafter the field is disaggregated into a hierarchy of fields, where each level in the cascade represents features in the rain field over a limited range of scales. The rate of change of the field at each level is estimated and used to introduce correlated noise at each level which increases in proportion with lead time, leading to pure stochastic noise after the largest scales in the field have evolved through their lifetime. The forecast probability distribution is estimated by generating 100 forecast sequences, each starting with the same observed rain field and for 10 to 120 minute lead times.