In past decades, there has been much scientific effort dedicated to the development of models for simulation and prediction of nitrate concentrations in groundwaters, but producing truly predictive models remains a major challenge. A time-series model, based on long term variations in nitrate fertiliser applications and average rainfall, was calibrated against measured concentrations from five aquifers in the River Frome catchment for the period spanning from the mid- 1970’s to 2003. The model was then used to “blind” predict nitrate concentrations for the period 2003-2008. To our knowledge, this represents the first “blind” test of a model for predicting nitrate concentrations in aquifers. It was found that relatively simple time series models could explain and predict a significant proportion of the variation in nitrate concentrations in these aquifers (R2: 0.6-0.9 and mean absolute prediction errors 4.2 – 8.0 %). The study highlighted some important limitations and uncertainties in this, and other modelling approaches, in particular regarding long term nitrate fertiliser application data. In three of the five aquifers (Hooke, Empool, Eagle Lodge), once seasonal variations were accounted for, there was a recent change in the generally upward historical trend in nitrate concentrations. This may be an early indication of a response to levelling-off (and declining) fertiliser application rates since the 1980’s. There was no clear indication of trend change at the Forston and Winterbourne Abbas sites, nor in the trend of nitrate concentration in the River Frome itself from 1965-2008.