Algorithms and Applied Econometrics in the Digital Economy

Abstract

This dissertation contains three papers that together study various aspects of the digital economy. Chapter 1 is entitled "The Impact of Financial Assistance on Online Learner Outcomes" and is an evaluation of the financial assistance program at edX, an online learning platform. I use Regression Discontinuity Design to evaluate the financial assistance program's effect on learner completion and certificate rates. I find large effects on both outcomes, even though it was free for learners to complete and pass a course during the period under study. These results indicate that learners value the signal of an edX verified certificate. To quantify the certificate's signaling value for the applicant subpopulation, I estimate a distribution of applicant willingness-to-pay. Lastly, I provide descriptive statistics for the applicant subpopulation, showing that financial assistance applicants are much more engaged than the average platform learner and also disproportionately live in countries with low or medium UN Human Development Index ratings. Chapter 2 is entitled "Nowcasting Trends in the US Housing Market." In this chapter, I extend a state-of-the-art flu tracking algorithm called Auto-Regressive with GOogle search as exogenous variables (ARGO) to nowcast trends in the US housing market. I show that ARGO improves nowcasting performance of housing market indicators at the state level and that ARGO with Zillow clickstream measures improves nowcasting at both the state and zip code levels. These results provide evidence that ARGO is a robust model that is applicable to economic domains and that clickstream data is a valuable source of information when used in a penalized model that avoids overfitting and is trained on a sliding window to capture changing usage patterns. To further understand the potential relevance of ARGO to economic nowcasting questions, I present preliminary evidence of ARGO's performance on nowcasting macroeconomic indicators and show that it performs reasonably well during times of economic turbulence by looking at how it would have performed during the Great Recession. Chapter 3 is entitled "Demand Learning and Dynamic Pricing for Varying Assortments" and is co-authored with my advisor Kris Ferreira of Harvard Business School. In this chapter, we develop a demand learning and dynamic pricing algorithm for a discrete choice setting with frequently varying assortments, where products are characterized by observable attributes and demand can be described by a multinomial logit (MNL) choice model. Our algorithm follows a learn-then-earn approach to deal with the well-known exploration-exploitation tradeoff. We increase the speed of learning during the initial learning phase by introducing methods from Conjoint Analysis to dynamic pricing. We evaluate our algorithm in a 90-day field experiment with an e-commerce company and find that, relative to the company's baseline pricing policies, our algorithm led to a significant increase in revenue over the 90-day period. We measure the treatment effects using synthetic controls and quantify the probability of observing the treatment effects using randomization inference with Fisher's exact test.