The Persistence of Discrimination Under Competition: A Dynamic Search Model with Endogenous Beliefs Working paper -comments welcome

Does competition eliminate discrimination when employer beliefs about minority productivity are formed through endogenous learning? Not necessarily. I embed Bayesian learning from own-firm hires in a competitive search model with free entry. The discriminatory steady state is the unique equilibrium and is globally stable: non-discriminating entrants face the same informational frictions as incumbents, which bounds their profit advantage and prevents entry from selecting the non-discriminatory outcome. A sharp threshold on signal precision separates two regimes: in the Walrasian limit (nesting Lepage 2024), any positive learning rate suffices for entry to close the gap; under search frictions with positive entry costs, even very precise signals are insufficient. A Shapley decomposition attributes most of the welfare loss to taste discrimination, but it is the information friction that forecloses Beckerian competitive convergence. A temporary hiring mandate generates persistent belief compression with a half-life governed by informational turnover (approximately 7 years), producing opposite predictions from models where mandates operate through biased evaluation criteria or worker investment distortions. Calibrated to U.S. labor market data, the model generates group-B expected income gaps of 6-15%.