Affective processes play an important role in determining human behavior in auctions. While previous research has shown that physiological measurements provide insights into these processes, it remains unclear which of the many features that can be computed from physiological data are particularly useful in predicting human behavior. Identifying these features is important for gaining a better understanding of affective processes in electronic auctions and for building biofeedback systems. In this study, we propose a new approach to identify physiological features for predicting auction behavior. We apply an Evolutionary Algorithm in combination with either the Multiple Linear Regression or Artificial Neural Network models to select physiological features and assess their predictive power. To test the approach, we use a unique dataset of participants' auction decisions and their synchronously recorded electrocardiography data. Our results show that the approach is able to identify subsets of physiological features that consistently outperform other physiological features.