This paper introduces a rule system to predict first-day returns of initial public offerings based on the structure of the offerings. The solution is based on a genetic algorithm using a Michigan approach. The performance of the system is assessed comparing it to a set of widely used machine learning algorithms. The results suggest that this approach offers significant advantages on two fronts: predictive performance and robustness to outlier patterns. The importance of the latter should be emphasized as the results in this domain are very sensitive to their presence.