The present study examined changes in the prediction of fear and panic in a clinical sample of patients (N = 25) meeting DSM-III-R criteria for panic disorder with agoraphobia (PDA). Data were collected for approx. 2000 trials of in vivo exposure. As expected, PDA patients displayed a bias for overpredicting both the level of fear and the likelihood of panic during an exposure trial. This overprediction bias was evidenced across several domains including heights, transportation and social situations. Although patients learned to make more accurate predictions within an exposure session, the level of overprediction remained relatively stable after the third trial within a session. Changes in fear overprediction differed across fear domains. Patients showed significant reductions in overprediction during exposure to transportation and social situations, but failed to show reductions in overprediction during exposure to heights. Unexpectedly, patients did not show increased prediction accuracy across sessions. These findings concur with earlier laboratory studies indicating that anxiety patients show an overprediction bias for panic and fear which decreases with practice. However, our findings indicate that the overprediction bias does not remit even after significant practice. The persistence of the overprediction bias is discussed within an evolutionary context.