Several recent studies have applied neural network models to the analysis and prediction of time series behaviour of solar activity. These studies have in general used overall measures of error magnitude to rate prediction accuracy. In this study, we concentrate on prediction of solar activity and demonstrate the tendency of neural networks to generate delayed predictions of specific features in the data. The necessity of recognising delayed predictions is discussed and a new training algorithm, based on a modified approach combining a genetic algorithm with back-propagation of errors, is proposed to counteract the problem.