The need to subject neural regression models to residual diagnostic tests is justified. A generalisation of the Durbin-Watson statistic is introduced for neural network regression models. The distribution of the statistic is estimated using approximation and exact methods (Imhof algorithm).