This paper addresses current inconsistencies in methodological approaches for neural network modelling of suspended sediment. An expansion in the number of case studies being published over the last decade has yet to result in agreed guidelines on whether suspended sediment load or concentration should be modelled, and whether log-transformation of data is either necessary or potentially beneficial. This contrasts with the well-recognized guidelines that direct traditional sediment rating curve studies. The paper reports a comprehensive set of single-input single-output neural network suspended sediment modelling experiments performed on two catchments in Puerto Rico. It examines the impact of internal complexity, input variable choice and data transformation on the form, consistency and physical rationality of model outputs, the existence of localized overfitting and the usefulness of global performance metrics. Sound guidance on whether to model sediment load or concentration, and whether to model log-transformed data is provided. Copyright © 2011 John Wiley & Sons, Ltd.