How Many Highly Underpriced IPOs Are Identifiable Ex Ante? Evidence from Out-of-Sample Predictions Using Machine Learning

We use machine learning techniques to conduct out-of-sample predictions of the underpricing of U.S. initial public offerings (IPOs) from 1990 to 2019. Using predicted underpricing based on ex ante information to sort the IPOs into 10 groups, we find that the underpricing averages for the top and bottom groups are 61.9% and 7.2%, respectively. Although pre-pricing public information is predictive, variables based on offering terms have primary predictive power. Our findings suggest that IPO underpricing is highly predictable, potentially allowing underwriters to reward their favorite clients with prodigious underpricing that cannot be justified by risk premia or information asymmetry-based theories