This paper introduces a comparison of one linear and two nonlinear one-step-ahead predictive models that were used to describe the relationship between human emotional signals (excitement, frustration, and engagement/boredom) and virtual dynamic stimulus (virtual 3D face with changing distance-between-eyes). An input–output model building method is proposed that allows building a stable model with the smallest output prediction error. Validation was performed using the recorded signals of four volunteers. Validation results of the models showed that all three models predict emotional signals in relatively high prediction accuracy.