A stochastic system for short-term spring flood forecasting for the Boyne River near Carman, Manitoba, has been developed. The system comprises two models. The upstream model consists of a transfer function noise algorithm which forecasts 6-h mean streamflow up to 36 h ahead for a site near Stephenfield, upstream of Carman. The inputs to this model are the recorded discharge of the Boyne near Stephenfield, temperature and precipitation during the melt period, cumulative winter precipitation, and an antecedent precipitation index for the previous summer and fall seasons. The downstream model is a transfer function noise model which uses previous discharges near Carman and recorded and forecast discharges near Stephenfield to forecast streamflow near Carman. Water-level forecasts near Carman are estimated from the forecast discharges with a stage–discharge relation for the site. Separate data sets were used to calibrate and verify the system. The system produced good forecasts for lead times up to 2 days. Key words: real time, floods, forecasting, snowmelt, simulation, transfer function noise modelling, antecedent precipitation index.