Delays in building contract execution can be predicted utilizing prognostic features of artificial neural networks (ANN). Trueness of prognostics highly depends of ANN topology. In order to choose the best one it is necessary to check which gives highest trueness. Trueness of ex-post prognostics can be calculated with use of traditional, statistic measures giving deviations of predicted values fromoriginal values for the test set of data, which is not involved in teaching process of ANN. It is proposed to measure the trueness by the width of defined in the paper bracket of accuracy for a given level of accuracy. The single prognosis is recognized as accurate if the predicted total time of contract execution is not shorter then 0,95 of the real one, and not longer then Lg of real total time of contract execution. Lg was defined as anupper limit of the accuracy bracket <0,95; Lg >. Evaluating trueness of two different ANN with the same accuracy level, the better one gives anupper limit Lg closer to 1. Then accuracy bracket will be more narrow.

Keywords: prognosis delays, artificial neural networks ANN.