Humans account optimally, or nearly so, for visuomotor variability (Trommershäuser et al., JOSA A 20, 2003) during movement under risk, even when variability is artificially increased (Trommershäuser et al., J Neurosci 25, 2005). How do humans estimate task-related variability in a dynamic, volatile environment in which variability changes? Methods: Subjects pointed rapidly at a target (a tall, green rectangle) on a visual touchscreen display. A small white circle indicated where the finger landed, and a blue circle was also displayed, randomly displaced horizontally from the white square. Displacements were drawn from a zero-mean normal distribution whose standard deviation remained constant for a sequence of trials (random epoch lengths: 75-150 trials) then suddenly changed to a new value (3.7-18.4 mm). Instructions indicated that outcome variability followed such a random sample-and-hold trajectory. On 2/3 of trials, an overlapping, horizontally displaced, red penalty rectangle was also displayed. Subjects won a point (4¢) if the blue square landed on the target, but lost 5 points if it landed on the penalty region. Observers should thus aim further from the penalty area when displacements are larger. Slow movements ([[gt]] 400 ms) were penalized 10 points. Results: 5 of 6 subjects took changing variability into account in planning movement. Actual movement endpoints (where the finger landed) were significantly correlated (p _{Conclusion: Subjects dynamically estimate movement outcome variability over a windowed running average of previous outcomes.}