The likely large-scale deployment of autonomous vehicle (AV) technology in the near future has the potential to fundamentally change the transportation landscape. Due to the high cost of AV hardware, the most likely path to widespread AV use is via platforms that can sustain high utilization, such as ride-hailing and delivery services. In this paper, we consider four potential operational models to commercialize AVs, which we model as a supply chain game between a platform, an AV supplier, and human drivers that join as individual contractors (ICs). Our operational models include (1) an open platform that outsources the high capital burden of AVs by allowing the AV supplier and human drivers to bring their own vehicles into the system, (2) an AV-only platform that is operated independently by the AV supplier, (3) a platform that sources AVs from the supplier through leasing contracts, and (4) an integrated supply chain in which the same entity operates the platform and supplies the AVs. We use (4) as a benchmark to measure the performances of the other models.

Platforms facilitating many-to-many matches in two-sided markets have become ubiquitous across industries ranging from professional services to dating. Differently from standard (one-sided) markets where consumers choose goods or services, in two-sided markets, both sides have preferences. Since these preferences can often be hard to describe, centralized matching is difficult to implement. The alternative option is for the platform to operate in a decentralized fashion, leaving the agents from both sides "free to find each other". While easier to implement, the downside of decentralized systems is that inefficiencies driven by congestion are likely to arise. In the present paper, we are primarily interested in understanding the power of "detail-free" levers that decentralized platforms can leverage to improve market outcomes. In particular, we focus on the lever of information design through competition signaling, where the platform discloses how much competition currently exists for a given supply unit. Signaling that there is competition for a supply unit may reduce the value of that unit but may also redirect the demand's attention to alternative supply units, potentially increasing the value for the platform. To quantify the trade-off at play and tackle the question above, we focus on a specific labor platform and the submarket of cleaning services to answer this question empirically. We partnered with the largest service labor marketplace in Latin America, which operates as follows. Service providers (agents) join the platform to purchase nonexclusive leads for jobs posted by supply-side customers. When they purchase a lead, they are not guaranteed to get the job, but simply purchase the contact information of the customer in order to apply for the job. A key characteristic of this market is the possible congestion on the lead side. In the context of such a platform, to understand the impact of any lever on market outcomes, it is fundamental to first understand how agents make their lead purchasing decisions and, in particular, how they take competition into account when making such decisions. We propose a structural model in which agents use a prediction function to forecast how much competition they may face. We show that if agents are strategic, a natural concept of equilibrium arises. By leveraging the platforms' data and an quasi-experiment, we estimate the structural parameters in the model. We find that agents react strongly to observed competition and predictions of future competition. We then conduct counterfactual analysis to study the impact of signaling competition. Our findings show that it is a powerful lever to improve market outcomes in this market. Signaling competition improves (decreases) congestion, and it also improves (increases) the probability that a lead will receive at least one applicant. Furthermore, displaying competition leads to an increase in overall leads purchased.

Marketplace companies rely heavily on experimentation when making changes to the design or operation of their platforms. A fundamental challenge in marketplace experimentation is dealing with interference. For instance, consider a ride-hailing platform experimenting with a demand-side price discount. The platform performs a randomized control trial (RCT), or A/B test, where some demand units are offered the discounted price while others are offered the undiscounted price. Because the treated units are more likely to book rides as a result of the discount, they reduce the total supply available to all demand-side units, including the control units. This interference between treatment and control units causes the Stable Unit Treatment Value Assumption (SUTVA) to fail, and consequently induces bias in the standard estimator used to evaluate the value generated by the treatment.

We study the problem of online contextual optimization where, at each period, instead of observing the loss, we observe, after-the-fact, the optimal action an oracle with full knowledge of the objective function would have taken. At each period, the decision-maker has access to a new set of feasible actions to select from and to a new contextual function that affects that period’s loss function. We aim to minimize regret, which is defined as the difference between our losses and the ones incurred by an all-knowing oracle. We obtain the first regret bound for this problem that is logarithmic in the time horizon. Our results are derived through the development and analysis of a novel algorithmic structure that leverages the underlying geometry of the problem.

We combine theory and empirics to (i) show that some buyers in online advertising markets are financially constrained and (ii) demonstrate how to design auctions that take into account such financial constraints. We use data from a field experiment where reserve prices were randomized on Google’s advertising exchange (AdX). We find that, contrary to the predictions of classical auction theory, a significant set of buyers lowers their bids when reserve prices go up. We show that this behavior can be explained if we assume buyers have constraints on their minimum return on investment (ROI). We proceed to design auctions for ROI-constrained buyers. We show that optimal auctions for symmetric ROI-constrained buyers are either second-price auctions with reduced reserve prices or subsidized second-price auctions. For asymmetric buyers, the optimal auction involves a modification of virtual values. Going back to the data, we show that using ROI-aware optimal auctions can lead to large revenue gains and large welfare gains for buyers.

We build a model to study the implications of utilization-based minimum earning regulations of the kind recently enacted by New York City for its ride-hailing providers. We identify the precise conditions under which a utilization-based minimum earnings rule causes marketplace instability, where stability is defined as the ability of platforms to keep wages bounded while maintaining the current flexible (free-entry) work model. We also calibrate our model using publicly available data, showing the limited power of the law to increase earnings within an open marketplace. We argue that affected ride-hailing companies might respond to the law by reducing driver flexibility.

We consider a firm that sells products that arrive over time to a buyer. We study this problem under a notion we call positive commitment, where the seller is allowed to make binding positive promises to the buyer about items arriving in the future, but is not allowed to commit not to make further offers to the buyer in the future. We model this problem as a dynamic game where the seller chooses a mechanism at each period subject to a sequential rationality constraint, and characterize the perfect Bayesian equilibrium of this dynamic game. We prove the equilibrium is efficient and that the seller’s revenue is a function of the buyer’s ex ante utility under a no commitment model. In particular, all goods are sold in advance to the buyer at what we call the positive commitment price.

We consider a multidimensional search problem that is motivated by questions in contextual decision-making, such as dynamic pricing and personalized medicine. Nature selects a state from a d-dimensional unit ball and then generates a sequence of d-dimensional directions. We are given access to the directions, but not access to the state. After receiving a direction, we have to guess the value of the dot product between the state and the direction. Our goal is to minimize the number of times when our guess is more than ε away from the true answer. We construct a polynomial time algorithm that we call Projected Volume achieving regret O(dlog(d/ε)), which is optimal up to a logd factor. The algorithm combines a volume cutting strategy with a new geometric technique that we call cylindrification.

We consider the problem faced by a firm that receives highly differentiated products in an online fashion and needs to price them in order to sell them to its customer base. Products are described by vectors of features and the market value of each product is linear in the values of the features. The firm does not initially know the values of the different features, but it can learn the values of the features based on whether products were sold at the posted prices in the past. This model is motivated by a question in online advertising, where impressions arrive over time and can be described by vectors of features. We first consider a multi-dimensional version of binary search over polyhedral sets, and show that it has exponential worst-case regret. We then propose a modification of the prior algorithm where uncertainty sets are replaced by their Lowner-John ellipsoids. We show that this algorithm has a worst-case regret that is quadratic in the dimensionality of the feature space and logarithmic in the time horizon.

No abstract available.

We study the perfect Bayesian equilibria of a model of social learning in networks where agents learn about an unknown state of the world by observing the actions of their neighbors. The network topology is drawn from an arbitrary distribution; contrary to prior models in the literature, two agents' sets of neighbors are not assumed to be independent. This extension allows us to capture real-world network phenomena, such as clustering and assortativity, and to consider the performance of social learning in widely used models of social networks, such as preferential attachment models. Since agents only observe the realization of their own neighborhoods, two agents could have vastly different beliefs about the overall network structure, and they may disagree over who is well-informed or well-connected. We call this phenomenon aggregate network uncertainty. This greatly alters our understanding of social learning dynamics and whether networks successfully aggregate dispersed information. Past literature has focused on herding outcomes as the key inefficiency of social learning, but we find that more severe inefficiencies can occur. In addition to the traditional metric of information aggregation, we introduce a second, weaker metric motivated by the notion of an expert, which we define as an outside agent whose private information is at least as strong as that of any other agent in the network. Learning is successful by our metric if all agents perform at least as well as an expert in the limit as society grows. Without aggregate network uncertainty, information aggregation is successful by this metric if and only if a connectivity condition holds. Herding outcomes can only occur once social learning has met our metric of success. With aggregate network uncertainty, there are several ways learning fails to reach even this weaker metric. Our main positive result is a characterization of sufficient conditions for successful information aggregation in the presence of aggregate network uncertainty. We show that information aggregation is successful whenever agents can identify well-connected neighbors with low distortion. Since neighborhoods are correlated, observing an agent may alter the informativeness of that agent's decision; distortion measures the extent to which this phenomenon occurs. We also demonstrate that in special cases our conditions are both necessary and sufficient for successful learning. We use this characterization to show that the popular preferential attachment network model successfully aggregates information.