In motor control, prediction of future events is vital for overcoming sensory-motor processing delays and facilitating rapid and accurate responses in a dynamic environment. In human ocular pursuit this is so pervasive that prediction of future target motion cannot easily be eliminated by randomizing stimulus parameters. We investigated the prediction of temporally randomized events during pursuit of alternating constant-velocity (ramp) stimuli in which the timing of direction changes varied unpredictably over a given range. Responses were not reactive; instead, smooth eye velocity began to decelerate in anticipation of each target reversal. In the first experiment, using a continuous-motion stimulus, we found that the time at which this occurred was relatively constant regardless of ramp duration, but increased as mean ramp duration of the range increased. Regression analysis revealed a quantitative association between deceleration timing and the previous two or three ramp durations in a trial, suggesting that recent stimulus history was used to create a running average of anticipatory timing. In the second experiment, we used discrete motion stimuli, with intervening periods of fixation, which allowed both target velocity and reversal timing to be varied, thereby decoupling ramp duration and displacement. This enabled us to confirm that the timing of anticipatory deceleration was based on the history of timing, rather than displacement, within the stimulus. We conclude that this strategy is used to minimize error amid temporal uncertainty, while simultaneously overcoming inherent delays in visuomotor processing.