Abstract
Based on three case studies, the impact of sample size and sample randomness on the predictive accuracy of multilayer perceptrons (MLP) is investigated. The {MLP} prove to be useful for classification problems. Although they are dependent on the sample size and the non-linearity of the underlying problem, they achieve predictions superior to the classical methods. A so-called saturation curve describes the dependency of the network performance on the sample size. This function enables the user to evaluate the achieved network performance and the usefulness of additional data. For reliable and generalizable results, the calculation of prediction intervals for the network is essential. It is demonstrated that the network leads to narrower confidence intervals of the performance measures in comparison to classical methods even for small sample sizes. The experiments show the validity of the law, for even relatively small sample sizes, that the standard error of the hit ratio decreases by one over the square root of the sample size. Therefore, the suggestion is to estimate the standard error for a given sample size by randomly drawing smaller sample sizes, and then rescaling the standard error accordingly.