Overconfidence is widely viewed as the costly belief bias; underconfidence is comparatively benign. We show this welfare ranking inverts in dynamic settings where agents select into feedback environments. Overconfident agents enter informative environments and self-correct; underconfident agents avoid them and remain miscalibrated, even under Bayesian updating. In our experiment, all participants make a manager choice, but only a randomly selected half have that choice implemented; the other half are randomly assigned a manager. Under choice, overconfident participants achieve substantially greater bias reduction than underconfident participants. Random assignment fully eliminates this advantage. The asymmetry is compositional---differential exposure through selection, not differential responsiveness.
You may participate online through the following ZOOM credentials
https://uoc-gr.zoom.us/j/88177475079?pwd=ik9vJRdyknP8xneS5NumlqgQdnplaz.1
Meeting ID: 881 7747 5079
Passcode: 941419
