We examine the use of feed forward neural networks in the long term (i.e., years ahead) prediction of sunspot number. First, we briefly review the history of the time series and also some previous attempts to predict it. We outline our neural network method and discuss how the reliability of the data affects training. We conclude that earlier data should not be used to train neural networks that are intended to make predictions at the current epoch. We then use this understanding of the data in training neural networks, testing many different configurations to see which provides the best 1--6 year ahead prediction accuracies. By looking at the distribution of residuals, an estimate of the uncertainty is placed on the best networks' predictions. According to our predictions of yearly sunspot number, the maximum of cycle 23 will occur in the year 2001 and will have an annual mean sunspot number of 130 with an uncertainty of ¿30--80% confidence. Finally, we discuss our result in relation to others and comment on how neural networks may be used in future work. ¿ 1998 American Geophysical Union