We have coupled a climate model (ECBilt-CLIO-VECODE) and a hydrological model (STREAM) offline to simulate palaeodischarge of nineteen rivers (Amazon, Congo, Danube, Ganges, Krishna, Lena, Mackenzie, Mekong, Meuse, Mississippi, Murray-Darling, Nile, Oder, Rhine, Sacramento-San Joaquin, Syr Darya, Volga, Volta, Zambezi) for three time-slices: Early Holocene (9000-8650 BP), Mid-Holocene (6200-5850 BP) and Recent (1750-2000 AD). To evaluate the model's skill in retrodicting broad changes in mean palaeodischarge we have compared the model results with palaeodischarge estimates from multi-proxy records. We have compared the general trends inferred from the proxy data with statistical differences in modelled discharge between the three periods, thereby developing a technique to assess the level of agreement between the model and proxy data. The quality of the proxy data for each basin has been classed as good, reasonable or low. Of the model runs for which the proxy data were good or reasonable, 72% were in good agreement with the proxy data, and 92% were in at least reasonable agreement. We conclude that the coupled climate-hydrological model performs well in simulating mean discharge in the time-slices studied. The discharge trends inferred from the proxy and model data closely follow latitudinal and seasonal variations in insolation over the Holocene. For a number of basins for which agreement was not good we have identified specific mechanisms which could be responsible for the discrepancy, primarily the absence of the Laurentide ice sheet in our model. In order to use the model in an operational sense within water management studies it would be useful to use a higher spatial resolution and a daily time-step. © 2006 Elsevier B.V. All rights reserved.