The impact of long-lasting non-point emissions on groundwater and streamwater in remote watersheds has been studied at numerous sites. In spite of substantially decreasing emissions in the last decade, recovery has not yet been observed in all cases. This trend might be masked by the considerable short-term variability of the chemical hydrographs. In this study, artificial neural networks are applied to investigate the SO 4 dynamics in the runoff of a small forested catchment susceptible to SO 4 deposition. Empirical models are fitted to the short-term dynamics at a time step of one day. About 75% of the variance of the SO 4 data is explained by the instantaneous discharge, short-term history of discharge and the moving average of SO 4 concentration in throughfall. In contrast, neither air temperature as an indicator for biological activity nor a snowmelt indicator based on the temperature sum increase the performance of the model. The model is used to investigate long-term trends in sub-regions of the phase space spanned by the identified input variables. According to the model, decreasing emissions have a significant effect on runoff SO 4 concentration only during the first severe storms at the end of the vegetation period. This suggests to focus on these events as indicators for recovery of the topsoil layers.