An algorithm of single component prediction based on backward error propagation is proposed, in which only one component concentration in a multivariate system is predicted each time. The algorithm was compared with a multiple component prediction model. In general, the predictive accuracy of the single component prediction algorithm was superior to that of the multiple component prediction model. The effects of overfitting, standard samples and model parameters on the predictive accuracy were also examined. © 1998 John Wiley & Sons, Ltd.