Supervised learning methods have been successfully used to build classifiers for the identification of promoter regions. The classifier is often built from a dataset that has examples of promoter (positive) and non-promoter (negative) regions. Thus, a careful selection of the data used for constructing and evaluating a promoter finding algorithm is a very important issue. In this context, experimentally known promoter regions can be safely assumed to be positive training instances. In contrast, since definite knowledge whether a given region represents a non-promoter is not generally available, negative instances are not straightforward to be obtained. To make the problem more complex, for the case of promoter, there is not a unique definition of what a negative instance is. As a consequence, depending on which definition of non-promoter region one assumed to build the data, such a choice could affect significantly the performance of the classifier and/or yield a biased estimate of the performance. We present an empirical study of the effect of this kind of problem for promoter prediction in E. coli. As far as we are concerned, up to now, there is no such a kind of study for the context of prokaryotic promoter prediction.