Research

Abstract: Peer-to-peer (P2P) reservation platforms often struggle to manage advance-booking customers. Because supply is self-scheduled, hosts' and customers' incentives may be misaligned. When platforms set high prices, the asset owners (hosts) would commit their assets early—i.e., list in advance; however, it may deter advance-booking customers from reserving. Conversely, setting lower prices would discourage hosts from listing early. Moreover, a host’s decision to list also depends on matching probability, which is governed by the platform’s design and search frictions--- that is,  its matching efficiency. We examine when and why advance-booking customers face a supply shortfall and how the platform's matching efficiency contributes to this outcome. We develop a two-period game-theoretic model in which the platform sets prices dynamically, and hosts choose when to list their assets. The model incorporates endogenous matching probabilities to capture real-world search frictions. To validate our theoretical insights, we analyze data from a major car-sharing platform, focusing on host listing decisions and their impact on service availability. We characterize the hosts’ optimal listing strategies in response to the platform's prices and identify three possible outcomes: no commitment, partial commitment, and full commitment. We show that when the demand imbalance between advance-booking and just-in-time customers exceeds a threshold, the platform prefers to induce a breakdown in the advance-booking market. Counterintuitively, this threshold decreases as the platform’s matching efficiency increases, meaning more efficient platforms are more prone to such breakdowns. Peer-to-peer reservation platforms should be cautious when choosing revenue-maximizing prices during severe demand imbalances, as it can disrupt advance-booking markets. Beyond pricing, a platform designer can use matching efficiency as a strategic lever to shape outcomes. By improving matching for early listers—through visibility boosts or recommendation algorithms—platforms can better align host incentives, reduce breakdowns, and enhance service across customer segments.

Abstract: Large language models (LLMs) are reshaping how people seek and produce technical knowledge. To evaluate the nature of knowledge where human expertise remains essential, we find a natural testbed in the largest technical Question and Answer community, StackOverflow. Each question asked in the community carries tags that mark the knowledge domains it draws upon. We term the unique combinations of these knowledge domains (i.e., unique sets of tags) a "specialization'' that characterizes the skill set needed to answer that question. Following ChatGPT's public release in November 2022, total question volume on StackOverflow fell sharply---a widely noticed trend. Yet beneath this decline, we discovered a notable trend: the proportion of novel questions that require stitching together knowledge domains in unprecedented ways rose significantly. More interestingly, we find that this surge in novelty emerges mainly from new recombinations of existing domains rather than new domains introduced by LLMs. This shift has also resulted in the reorganization of the community's knowledge structure. In particular, modeling the knowledge in the community as a network of co-occurring domains, we find that popular knowledge domains, i.e., historically well-discussed domains, have weakened significantly, as activity migrates toward niche domains. Consequently, the community network becomes more fragmented with a weaker core. Our findings suggest that LLMs substitute for standardized queries, while novel problems that require creative integration of knowledge domains still need human expertise. Such selective substitution of questions has profound implications for the sustainability of online knowledge communities and for the reliability of training data that future AI systems will depend on.

Abstract: Online Q&A communities have transformed customer support by decentralizing knowledge exchange. However, their success may differ, illustrated by Stack Overflow's growth under strict moderation compared to Yahoo Answers' decline despite weaker moderation. This paper investigates how community design, particularly the cost of asking questions, shapes user participation and community survival. We develop a multistage stochastic game model where users with heterogeneous skill levels dynamically decide to join, leave, ask, and answer questions. Our framework explicitly incorporates the ``cost of asking" as the effort users must exert to meet community standards and avoid moderation penalties. We first characterize the unique steady-state equilibrium of user participation. Under this equilibrium, we establish conditions under which empirically observed core-periphery network structures are the only possible configuration, thereby providing a theoretical validation for existing empirical work. Counter to common intuition, our findings reveal that a community's survival critically depends on the joining of low-knowledge users, while it can persist even when high-knowledge users do not join. Finally, our analysis highlights a strategic trade-off for community designers between participation level and user retention. We demonstrate that the total number of active users in the community is non-monotonic in the cost of asking questions, revealing that increasing the cost of asking is not always detrimental and can, in fact, improve user satisfaction.

Abstract: Payment for results (PfR) funding approach, where donors reimburse the non-profit organization (NPO) based on outcomes, is being increasingly adopted in the non-profit sector. However, there is also concern expressed by many voluntary organizations that such a funding approach puts an undue financial burden on small NPOs and could actually be detrimental to social welfare. In this study, we build a theoretical framework to analyze PfR funding mechanisms. We use a sequential game to model the interaction between the donor and the NPO, with the donor as the first mover. This model captures how PfR funding is typically implemented in practice using social impact bonds (SIB), wherein social investors provide the upfront funding needed by the NPO to implement the project. The donor provides funding, based on the actual benefit delivered, at the end of the project and the investors are paid back using these funds. We find that higher targets set by the donor do not necessarily translate to higher expected utility or expected benefit delivered under PfR. When comparing the performance of PfR and traditional funding (TF) mechanisms, we find that the donor typically has a higher expected utility under the PfR mechanism when the probability of a negative outcome shock is either high or low, and is better off using the TF approach otherwise. When the donor’s opportunity cost of funding the project is high, the donor is better off using a PfR mechanism when her belief about the NPO having low efficiency is sufficiently high. Interestingly, we find that for a large range of parameter values there is a mismatch between the approach that gives a higher expected utility to the donor and the approach that maximizes the expected social benefit delivered. Our model and analysis suggest that the optimal funding approach, and the optimal target set under PfR, depend on the NPO’s financing cost from social investors and project outcome uncertainty.

Abstract: In a shared economy, users list their assets on the platform for extra earnings. However, in emerging markets, there may not be enough individuals who are willing to share assets. Some platforms nowadays offer asset financing options to ensure enough supply, wherein individuals can own the asset temporarily or permanently by committing to a recurring fee. These users can then list the asset on the platform and earn a share in the revenue. We study such contracts using a sequential game-theoretic model where the platform chooses the recurring fixed fee and revenue share. The individuals then decide to subscribe or not using rational expectations on their future payoffs. The platform can choose a contract with a high subscription fee and high revenue share or one with a low subscription fee and low revenue share. We study conditions under which the platform prefers to choose the former contract over the latter. Specifically, we study the effect of platform prices, the asset's market price, and the demand for the shared asset on the contract design. Our main result is the existence of a budget limit above which the platform will not exhaust its budget to get more subscribers. Interestingly, depending on the market conditions, this budget limit can be lower than that required to get subscribers to satisfy low demand levels. We further find that it is optimal for the platform to have subscribers who list their assets infrequently, even when the platform is supply-constrained. We use the data from a car rental platform in India that offers such subscription contracts to validate our results and support our modeling assumptions. Finally, we also compare the profits of a platform offering subscription contracts against a centralized platform and study if such contracts are always preferable to scale.

Abstract: Lenders in emerging markets face a critical operational challenge: dynamically allocating scarce repossession resources across heterogeneous borrowers whose payment behavior is uncertain and interdependent. Black-box machine learning offers predictive power but lacks interpretability, and deployment challenges in regulated financial operations, while traditional rules-based systems ignore learning opportunities. We develop an explainable AI framework that combines the strengths of both approaches for electric vehicle lending—a high-impact application where repossession decisions affect both financial sustainability and access to clean transportation. We formulate the problem as a budget-constrained restless multi-armed bandit with outcome-triggered observability: repossession costs and salvage values are heterogeneous, payment probabilities depend on market-level exposure variables, and the lending regime (platform-recommended vs. self-chosen price) is observed only for booked loans. Rather than deploying black-box reinforcement learning, we develop a fluid relaxation that yields a tractable linear program whose dual solution provides an interpretable priority index for repossession decisions. Using data from 12,347 loans, we show our hybrid approach achieves 17–23% improvement over traditional heuristics with provable competitive ratios (≥94% of fractional optimum), and faster computation. 

Abstract: Multi-brand cloud kitchens promise customers variety by co-locating multiple restaurant brands under one roof and enabling one-basket checkout across brands. Yet in a dataset of 6.24 million orders from 68 kitchens, multi-brand orders take substantially longer to fulfill and receive lower customer satisfaction ratings than single-brand orders. We show that this performance gap is primarily attributable to information architecture and demand steering rather than inherent coordination complexity. By restricting comparisons to orders placed via the platform channel and decomposing order types by the presence of cross-listed dishes—items listed on one brand's menu but prepared by another—we find that coordination penalties are load-activated and substantially higher when coordination is chef-led (cross-listed dishes visible only to the host chef) than synchronizer-led (explicit multi-brand orders visible to a dedicated kitchen officer). Brand ownership effects manifest through congestion asymmetries created by demand steering rather than contractual differences. Using a fork–join queueing model calibrated via Simulated Method of Moments, we simulate four counterfactual policies and find that upgrading information systems to display cross-listed dishes on all participating screens reduces fulfillment time by 2–5\% for multi-brand orders while improving single-brand performance—a Pareto improvement. Our findings demonstrate that in distributed coordination systems, information visibility is a first-order operational lever.