Accurate predictions of the future behavior of the solar activity cycle have been sought for many years. Several classes of prediction approach have been proposed, with many variations in each class, and have achieved varying degrees of success. However, considerable room for improvement still remains. Artificial neural network models enjoyed a resurgence in popularity as prediction tools during the late 1980s, as a consequence of the discovery of the back propagation of errors learning algorithm. Initial investigations have been carried out into their potential for predicting solar activity (e.g., Koons and Gorney, 1990; Williams, 1991; Macpherson, 1993a, b). In this paper, we investigate in detail the effect different neural network architectures and learning parameters have on the prediction accuracy of various networks trained on smoothed monthly sunspot and solar 10.7-cm flux data. The importance of obtaining the best generalization capability of a neural network is stressed. Prediction of the geomagnetic aa index is also considered. Finally, in order to validate the usefulness of this technique, the results are compared with a variant of the well-established McNish and Lincoln method (McNish and Lincoln, 1949) and are found to be superior in terms of prediction accuracy. ¿ American Geophysical Union 1995