I am an economist at the University of Pennsylvania's Wharton School. I specialize in economic theory and its applications to other areas of economics, to science, and to business. I am particularly interested in market design, selection markets, social science genetics, and how organizations can use online and offline experimentation to increase productivity.
I received my Ph.D. in Economics from Harvard University, advised by Alvin Roth, and was a 2016 Sloan Foundation Fellow.
Eduardo Azevedo
John M. Bendheim, W'40 and Thomas L. Bendheim,
WG'90 Professor of Business Economics and Public Policy
In industries such as airlines, steel, and truck rentals, firms have coordinated to raise prices or reduce capacities using public announcements such as earnings calls. As this type of coordinating practice is relatively unexplored, this study seeks to learn more about the collusive content in firms' public communications. To do so, we design a large language model (LLM) to look for such content and then apply it to almost 500,000 company transcripts. The LLM is very selective in that only .12% of the transcripts satisfy the criterion for collusive content. The LLM is reasonably effective at rediscovering previously documented episodes and identifies many new episodes. A human audit is conducted of the 301 transcripts with the highest LLM scores. The audit distills and organizes the content to identify systematic ways in which it is facilitating coordinated conduct. Messages fall into two broad bins. First, a firm states that the industry needs to raise prices, reduce supply, or reduce capacity. Second, a firm expresses how it has been less aggressive and calls on competitors to join them, either explicitly or implicitly by saying it is "acting as a leader", commending rival firms who have acted in a similar manner, or criticizing rival firms that have not. A second research objective is to develop a screening tool for competition agencies and plaintiff law firms. Towards that end, the human audit evaluates the performance of the LLM in substituting for human experts. The LLM is capable of identifying a wide variety of collusive content though there is a reasonably high false positive rate. Nevertheless, we explain how it can be an essential component of a screening protocol and show proof of concept with a case study of the Brazilian domestic airline industry.
Recent proposals to tax unrealized capital gains have sparked a debate about their impact on entrepreneurship. We show that accrual-based taxation creates two opposing effects: successful founders face greater dilution from advance tax payments, whereas unsuccessful founders receive tax credits that effectively provide insurance. Using comprehensive new data on U.S. venture capital deals, we find that founder returns remain extremely skewed, with 84% receiving zero exit value while the top 2% capture 80% of total value. Moving from current realization-based to full accrual-based taxation would reduce founder ownership at exit by 25% on average but would also increase the fraction receiving positive payoffs from 16% to 47% when tax credits are refunded. Embedding these distributions in a dynamic career choice model, we find that founders with moderate risk aversion prefer the current realization-based tax system, while more risk-averse founders prefer accrual-based taxation. We estimate that a 2% annual wealth tax has a similar impact on dilution as taxing unrealized capital gains, but produces limited risk-sharing benefits due to the absence of tax credits in case of down rounds.
The predictive power of genetic data has been increasing rapidly and is reaching levels of clinical utility for many diseases. Meanwhile, many jurisdictions have banned insurers from utilizing genetic information. This has led to concerns that further improvements in genetic prediction will lead to adverse selection. We make three contributions to this debate. First, we develop a method to measure the amount of selection in an insurance market where consumers have access to current genetic prediction technology. Second, we extend the method to estimate the amount of selection given expected improvements in genetic prediction technology. Third, using the UK Biobank dataset with nearly 500,000 genotyped individuals, we apply the method to the critical illness insurance market. We find that expected improvements in genetic prediction are likely to lead to unsustainably high levels of selection and thus threaten the viability of the market. We discuss policy implications.
We consider the standard moral hazard problem with limited liability. The first-order approach (FOA) is the main tool for its solution, but existing sufficient conditions for its validity are restrictive. Our main result shows that the FOA is broadly valid, as long as the agent's reservation utility is sufficiently high. In basic examples, the FOA is valid for almost any positive reservation wage. We establish existence and uniqueness of the optimal contract. We derive closed-form solutions with various functional forms. We show that optimal contracts are either linear or piecewise linear option contracts with log utility and output distributions in an exponential family with linear sufficient statistic (including Gaussian, exponential, binomial, geometric, and Gamma). We provide an algorithm for finding the optimal contracts both in the case where the FOA is valid and in the case where it is not at trivial computational cost.
