Reputation concerns in risky experimentation

We develop a general model, with the exponential bandit as a special case, in which high-ability agents are more likely to achieve early success but also learn faster that their project is not promising. These counteracting effects give rise to a signaling model in which the single-crossing condition fails but a double-crossing property holds. We characterize the unique D1 equilibrium under double-crossing condition, and show that it tends to produce pooling. Ability to identify good projects and ability to execute a good project have different implications for the equilibrium allocation. Our model also incorporates public news, which generates dynamic distortions.