There are two well known dimension of emotions, arousal and valence. In this research we focus on arousal. The PicDisengage dataset consists of 43 users who are shown 30 images for 32 trials per image (1.5 to 2.5 s per image) while their eye movements are recorded using a state-of-the-art eye tracker. For each image the users give a manual arousal rating between 1 and 9. We apply to this data a well known machine learning pipeline for eye tracking data that has been successfull for user prediction and other tasks. By applying this pipeline methods we achieve successful prediction of arousal ratings, as described below. When we do binary prediction of arousal ratings 1–4 versus 5–9 then we obtain an accuracy of 78%. When we do binary predictions of one arousal rating versus the rest of the ratings then we obtain accuracies always above 80% and in some cases as high as 96%. More precisely, for predicting the arousal ratings 5, 6, and 7 against the respective rests we obtain accuracies of around 82%, while for all other arousal ratings our accuracies are above 90%. The three major achievements of our research are: (i) it enables prediction from static images alone (whereas all prior studies on the top use video clips plus possibly other modalities), (ii) it requires very little data (instead of the full 32 trials our predictions achieve similar accuracies when using only 6 trials), and (iii) our predictions are made using unseen images (this corresponds to “cross stimulus prediction” which is an extremely challenging and largely unsolved issue within machine learning based eye tracking research).