Ron Azuma showed with user testing that even with true acceleration and angular rotation measurements, prediction starts making things worse after 60ms. (per Russ Taylor, citation needed)
Russ Taylor, personal conversation: The two things to look out for with the predictor are: (1) overshoot and bounce-back when the user stops moving their head (predicting too far ahead), and (2) Jittery behavior even when sitting still (too much noise in the tracking system for dead-reckoning prediction without a Kalman or other filter).
Rich Holloway's "For a standard task, you get about 1mm of position offset for every 1ms of tracker latency." (per Russ Taylor, citation needed)