The first thing I though was a bad synchronization, but the anticipation was 3 seconds, that is big enough.
The second possibility is the filtering process. It's weird and it's a common source of error. The interesting part is in page 21-22:
> Low consistency responders
> Given that there was better evidence for presentiment from the consistent responders, one wonders how the inconsistent responders performed?
> Examination of the raw data revealed that in most cases, the inconsistent responders were so labeled because one or two of their calm trials had exceptionally large within-trial variances (i.e., variance of the physiological measure from the moment of the stimulus to the end of the cool-down period). Because of this observation, as a post-hoc test we examined the mean presponse correlation for SCL for each inconsistent responder, after removing the one calm trial with the most extreme within-trial variance.
> Table 9 shows that the effect of removing this single trial from each of the total of 49 inconsistent responders (across both experiments), indicated as I* in the Table, dramatically changed the overall results.7
> All data from all subjects combined resulted in a nonsignificant zpre-r = 0.03, whereas removing the single highest-variance calm trial from each of the inconsistent subjects (leaving 98.5% of all data) resulted in a combined significant zpre-r = 2.99 for SCL.
So the problem is that they are filtering the people that made a random move just before seeing a calm image, but they are not filtering the people that did a random involuntary move just before seeing an emotional picture.