Learning and Overconfidence Bias: Evidence from US College Debaters

Overconfidence in one’s abilities does not seem to be tempered by experience, suggesting that overconfidence serves an evolutionary purpose. Compte and Postlewaite (2004) argue that if confidence improves performance, the welfare maximizing learning technology ignores some negative information and overconfidence is optimal. This theory suggests that peoples’ self-assessments should improve over time, converging to a biased level. We test this claim among university debaters, and find that while experienced debaters are no less biased than inexperienced debaters, their predictions are half as noisy around this biased level. Debaters, a group for whom confidence is critical for performance, learn over time; but they never learn that they are not as good as they think they are. We therefore provide some evidence of the origin for this widespread behavioral bias.