Perhaps the ease of dealing with the problem depends on the sampling rate of the gaze data, but you can try to first identify all saccades by velocity and/or acceleration criteria using two-point differentiation (or other methods), and determine, in relation to the direction in which the text the subject is reading advances, whether the gaze position after each identified saccade is ahead or behind the gaze position before the saccade.

This piece of software may be helpful: Saez de Urabain et al (2015) GraFIX: A semiautomatic approach for parsing low- and high-quality eye-tracking data, Behavior Research Methods

 http://link.springer.com/article/10.3758/s13428-014-0456-0?sa_campaign=email/event/articleAuthor/onlineFirst

Plot the data: X(for a given Y) vs time. That is the easiest way to identify regressions.

A higher sampling rate will yield more useful information