Research

We study a dynamic relationship in which a long-lived principal delegates experimentation to short-lived agents, each living for one period. Therefore, the principal has more incentives to experiment than agents. We model this as a contextual bandit problem, where the experimentation yields success that follows a state-dependent Poisson process. The state is drawn independently at the beginning of each period, and a new agent arrives each period with private knowledge about the state. The success rate is unknown to the principal and all agents in one of the states, and it can either be high or low. Players will learn it through experimentation. We focus specifically on stopping-time policies. Due to state transitions, we show that experimentation tends to follow a cyclical pattern—experiencing phases of interest, abandonment, and revival. Meanwhile, the conflict of interest leads to both over- and under-experimentation.

Market Design as Joint Signaling-Screening Mechanism