Pricing and Information Design for On-Demand Platforms
[Job Market Paper]
I study how to optimally design on-demand platforms. Formally, I consider the design of information and pricing mechanisms in an unobservable queue with strategic and heterogeneous customers that are privately informed.
First, I show that if the platform is forced to release information about the level of congestion, then—among all possible pricing mechanisms—the optimum can be implemented by a sequence of posted prices that depend on the level of congestion.
Second, I show that if the platform can jointly optimize among all possible pricing and information mechanisms, the optimum can be implemented by disclosing all information and a sequence of posted prices that depend on the level of congestion.
Third, I identify situations under which information design is profitable and provide a linear program that allows for the computation of the optimal policy.
Finally, I show that when consumers’ private information is multidimensional, the loss from disclosing the amount of congestion is arbitrarily large relative to the overall optimal mechanism.
The Amplification of Discrimination via Crowdsourcing in Online Platforms (with Shunya Noda)
It has been well documented that minority workers (e.g., people of color, women, people with disabilities) get a significantly worse reputation than majority workers in a variety of crowdsourcing platforms. In this paper, we develop a tractable dynamic model of a crowdsourcing platform, in which consumers review workers’ task performance. We show that if the platform discloses the full set of past reviews and a small fraction of consumers discriminate against minority workers (make biased reviews), then minority workers are discouraged from exerting effort since they cannot develop a sufficiently strong reputation to find doing so profitable. Our result suggests that poor rating and platform design may amplify discrimination and endogenously decrease the observed productivity of minority workers.
Informed Seller: Personalized Pricing versus Information Design (with Shota Ichihashi)
We study the problem of a seller that is privately informed about the quality of a product she is selling and knows from which distribution the consumer’s type is being drawn. The consumer cares about quality and has a privately known signal about his utility. First, we show that revealing all information about the quality of the product together with a sequence of personalized posted prices that depend on the quality of the product attains the seller’s overall optimal revenue. Second, we show that if the seller is forced to: (a) reveal information about the quality of the product and (b) employ anonymous pricing mechanism, then the optimum can be implemented by setting a sequence of posted prices with respect to the mixture distribution. Finally, we show that—if the seller is only forced to employ an anonymous mechanism—personalized information disclosure policies improve revenue.
Beliefs about the Future and Temporal Invariance (with Federico Weinschelbaum)
Thompson  identifies a non-essential transformation: "interchange of decision
nodes". The implication from this seminal work is that permuting the order of
simultaneous plays should be irrelevant for the equilibrium prediction. However, this
is not the case. Most equilibrium refinements that operate on the extensive-form imposing
structure on beliefs off-the-equilibrium path fail to satisfy this basic idea of
"temporal invariance". To address this issue, we propose a new equilibrium concept based on adequately
defined beliefs (i.e. beliefs that treat past and future actions in the same way) that satisfies
the following desirable properties: Bayesian Updating, Sequential Rationality,
Backward Induction, Forward Induction, and Temporal invariance.
Repeated Auctions: Simple Learning Rules Under Binary Priors
I consider a seller who repeatedly runs a sequence of second price auctions to sell identical nondurable goods to a set of buyers whose valuations for the products are drawn from a fixed distribution. At the beginning of the game, the seller does not know what the right distribution is, but may employ past histories to try to learn the optimal reserve prices.
In particular, I show that when buyers are myopic and do not take into account the seller’s learning, a greedy pricing (GP) rule leads to full and exponentially fast learning. However, when buyers are forward looking, GP may induce buyers to behave strategically and lead to incorrect learning. To circumvent this issue, I introduce a notion of approximate incentive-compatibility in conjunction with a random greedy pricing (RGP) rule, and show that this restores the previous learning results. Finally, I illustrate the connection between learning speed and the likelihood of implementing truth-telling as an equilibrium.