Research

Research interests

Substantive

• Experiential consumption, prosocial behavior, customer analytics and management 

Methodological

• Bayesian econometrics, discrete choices, machine learning

Projects

Customizing Bundles of Experiential Goods - An Application to Performing Arts Ticket Sales

with Fred Feinberg and Eric Schwartz

Revise and Resubmit at Journal of Marketing Research

Firms proffering experiential goods like performing arts and meal kit services increasingly allow consumers to curate the “bundle” they purchase, allowing enhanced flexibility and customization. This comes at the cost of greater consumer cognitive demand, which firms attempt to mitigate through pre-set “themed” packages, as opposed to proposing suitable bundles for each customer. To this end, we develop a model leveraging past custom bundle choices to suggest future ones, balancing utility for individual component items with “holistic” bundle-level qualities. The model is applied to create-your-own performing arts subscriptions, and overcomes two challenges ubiquitous for experiential goods – the need to make predictions for items never before encountered, and the vast number of potential bundles to be chosen from – via natural language processing and a suite of bundle-level metrics that align with human perceptions of within-bundle variation. The model embeds heterogeneous preferences in a hierarchical Bayes specification and, to tame curse-of-dimensionality, is estimated via a permutation-based Monte Carlo approach that integrates over a sequential factorization of the joint bundle likelihood. Results suggest that most consumers prefer bundles with a pronounced degree of topical cohesiveness; moreover, systematically incorporating this improves out-of-sample bundle choice predictions by 7.2% on a per-performance and 50.7% on a per-bundle basis. That model-proposed bundles have higher topical cohesiveness suggests that empirical bundle choices differ from simply assembling one’s most-preferred items, and can guide how experiential goods firms might leverage user histories to propose superior customized bundles.

Cross-channel Price and Inventory Optimization for Live Events: An Application to NFL Ticket Purchases

with Fred Feinberg and Pete Fader

In close collaboration with a National Football League (NFL) team and a firm that specializes in sports analytics, we investigate the purchase channel choices of ticket buyers with the goal of optimizing inventory and pricing policies across multiple channels. NFL tickets, like other professional sports leagues, are distributed through two types of channels: an official marketplace (i.e., primary channel) where the team directly distributes tickets and secondary channels, such as StubHub, where individuals and resale brokers sell their tickets. These sellers release their inventory at different times and price them dynamically, which, combined with the uniqueness of individual tickets (i.e., there is only one ticket for a specific seat for a specific game), makes availabilities and prices vary by when and where consumers choose to buy their tickets. This complexity makes ticket purchase for such events a nontrivial undertaking, especially if seeking a “block” of contiguous seats, one that requires extensive information gathering, search, and anticipation of trends on availability and pricing. In this paper, we study consumers’ choice of purchase channels, including how channel-specific observed price trajectories and availability, as well as individual past-dependence (e.g., inertia) affects channel choices. This in turn allows us to understand how different pricing and inventory management policies affect channel choices, as well as for policy optimization. To this end, we collected unique availability and pricing data on four major ticket exchange platforms, including the primary channel and three secondary ones. We combine this with the rich historical purchase data provided by the team, which includes seat-level transaction with channel-choice information. Our preliminary findings suggest that users have a strong past-dependence in their channel choice even after controlling for prices and the number of tickets that they seek, and that users are significantly more likely to use secondary channels as the group size (number of tickets that they buy together) grows larger.

Leveraging Gaussian Processes for Counterfactual Inference in One-Shot Field Experiments: Evidence from a Large-Scale Donation Drive 

with KeeYeun Lee and Fred Feinberg

"Appeals scales," sequences of suggested amounts presented to potential donors, are a key solicitation tool for charities. In this project, we investigate how altering appeals scales affects donation outcomes – both incidence and amount – using a fully orthogonalized field experiment  conducted by a large international charity in France. Two primary components of the appeals scales were manipulated: minimum and maximum suggested amounts (consequently, “range”) and the pace of increase (“slope”). The experiment is unusual in the sense that the randomization led to some respondents being asked to donate less than they had in the past, entailing the potential to hurt the sponsoring charity, but offering a unique opportunity to calibrate the effects of manipulating scale components at the individual-level. Measuring "causal" effects of scale manipulation requires that individual-level donation history be properly accounted for, so as to calculate counterfactual imputations for three key components: when, whether, and how much the user would have given in the absence of their specific experimental manipulation. To control for the timing and quantity of prior donations, we propose hierarchical multi-output Gaussian Processes for latent propensities over both donation incidence and amount. This approach allows us to fully leverage individuals’ history of donation – rather than summary metrics of donation histories – to measure the effects of scales with individual-level counterfactuals. This would allow charities to improve or potentially optimize their future appeals scales at the individual-level, including when to solicit donations at all. 

Predicting commitment: leveraging donation histories to incent "subscription" to charitable donation 

with Fred Feinberg and Jen Shang

Charity-donor relationships are analogous to firm-customer ones, broadly maintaining either a contractual or a non-contractual relationship. Specifically, in a contractual relationship with a charity, donors engage in subscription-like regular giving, i.e., a recurring gift consisting of a fixed amount at a regular interval (often monthly). By contrast, in a non-contractual relationship, one might give irregularly, with patterns that vary by timing and amount. However, charity-donor relationships can be quite flexible, as donors can and do give in both forms. But, more importantly, one can convert from a non-regular to a regular donor. We focus on this group of converters, given their outsized importance in providing a sizable stream of revenue (78% in our focal charity from the U.K.) and allowing charities to forward-plan by offering predictable cashflow. 

Specifically, we analyze the donation patterns of "converters", who commit to giving on a regular basis after some period of irregular giving, with an eye towards detecting and "nudging" those nearing conversion. Using a survival analysis framework, we find that there are detectable patterns – increases in relative frequency and decreases in recency – in the periods that lead to commitment to a regular giving. The next phase of the project will leverage hierarchical Gaussian Processes to capture individual trajectories with different length scales to account for the wide heterogeneity in the (calendar) time to commitment and to detect changes that can signal "readiness to convert". Optimization scenarios based on the model’s results can help charities design and target interventions to induce donors to convert.