Described herein are techniques for adjusting a prediction level and a throttle level, as frames are being rendered on a head-mounted display (HMD), based on an application's rendering performance. The prediction level is increased if a number of late frames, out of a past N rendered frames of (N being any suitable number), meets or exceeds a threshold number of late frames, which causes a compositor of the HMD to predict pose data of the HMD farther out into the future. The throttle level can be increased independently from, or in synchronization with, the increase in the prediction level to causes the compositor to throttle the frame rate of the application (e.g., to a fraction of the refresh rate of the HMD). The prediction level (and the throttle level, if at the same level) can be decreased if a particular number of consecutively-rendered frames finish rendering early.