It is known that arbitrary target accelerations along the horizontal generally are extrapolated much less accurately than target speed through a visual occlusion. The extent to which vertical accelerations can be extrapolated through an occlusion is much less understood. Here, we presented a virtual target rapidly descending on a blank screen with different motion laws. The target accelerated under gravity (1g), decelerated under reversed gravity (-1g), or moved at constant speed (0g). Probability of each type of acceleration differed across experiments: one acceleration at a time, or two to three different accelerations randomly intermingled could be presented. After a given viewing period, the target disappeared for a brief, variable period until arrival (occluded trials) or it remained visible throughout (visible trials). Subjects were asked to press a button when the target arrived at destination. We found that, in visible trials, the average performance with 1g targets could be better or worse than that with 0g targets depending on the acceleration probability, and both were always superior to the performance with -1g targets. By contrast, the average performance with 1g targets was always superior to that with 0g and -1g targets in occluded trials. Moreover, the response times of 1g trials tended to approach the ideal value with practice in occluded protocols. To gain insight into the mechanisms of extrapolation, we modeled the response timing based on different types of threshold models. We found that occlusion was accompanied by an adaptation of model parameters (threshold time and central processing time) in a direction that suggests a strategy oriented to the interception of 1g targets at the expense of the interception of the other types of tested targets. We argue that the prediction of occluded vertical motion may incorporate an expectation of gravity effects.