Information technology tools and resources at the UW
Creating New Science with Amazon Web Services
What if Airlines Had Not Merged?
UW-IT’s new partnership with Amazon Web Services (AWS) provides a platform for visiting PhD student Charlie Manzanares that is allowing him to do new science in the field of economics that exceeds what most economics researchers are typically able to do. One of Charlie’s research questions—what would airline travel and pricing look like today if major carriers had not been permitted to merge—may assist the United States Department of Justice (DOJ) in determining whether airlines have colluded in setting prices. In the spring of 2015, the DOJ announced that it is investigating collusion among U.S. airline carriers who may have restricted capacity unlawfully.
AWS has enabled Charlie to estimate dynamic models of airline competition at scale and quickly across the entire U.S. airline network, which is something scholars struggle with in the absence of cloud computing.
“If things go well, our model could be one of the first in the U.S. to provide antitrust authorities with the ability to estimate the impact of this type of collusion on consumer welfare, at scale, which might be useful to the DOJ as they develop their investigation.” Charlie will be presenting his work at UW’s Cloud Day on November 12.
Rich Data Sets
“Airline data is one of the best data sets we have in economics. It’s a remnant of the era when airlines were heavily regulated, which ended in 1978,” he explains. Under the pre-1978 system of federal regulation, the government set prices and schedules for airlines. When the government de-regulated the industry, they continued to require airlines to provide a sample of ticket prices and segment activity. The segment data set includes the number of passengers transported over each route flown by U.S. domestic carriers with a U.S. endpoint. The price data set represents a 10% sample of all airline tickets that have been purchased in the U.S. since 1993. The Department of Transportation collects and manages this extremely detailed data. The data set that Charlie works with begins in 2003.
The Role of Dynamic Game Theory
Building on an early collaboration with the UW eScience Institute and with guidance from the team at Amazon, Charlie has been working on a research program that focuses on developing statistical methods for studying dynamic games, which he then applied to airlines. This program represents the intersection of dynamic game theory, machine learning, and econometrics. Dynamic game theory considers variables that change over time that affect the way the game is played out by competitors.The goal is to figure out how each player would react as the game evolves. An airline game is realistically modeled using thousands of state variables. The complexity of this game generates a statistical and computational problem that has a dramatic name: “the curse of dimensionality.” The iterations of variables become too numerous for modern computing power to resolve.
The machine learning method uses the data to select a smaller number of state variables, which are the ones that it deems to be the most important for describing airline competition. Using a smaller number of state variables lowers the computational burden of simulating dynamic competition between airlines. (That task is quite large even with machine learning and certainly is far too big for a laptop—each iteration would take at least a couple of weeks.) Dynamic games between airlines are simulated using actual data as inputs in order to try to achieve the best model of airline competition. That in turn allows Charlie to start simulating the effect of recent mergers.
This work is particularly salient given the fact that in the past decade the U.S. airline industry has experienced the most dramatic merger activity in its history. Since 2005, eight major carriers have merged down to four: American, Delta, United/Continental, and Southwest. Those carriers account for 80% of all airline traffic. Charlie’s work is creating a model that shows what airline prices and flight networks would have looked like if those mergers had not been permitted. That model can be used to study, for example, whether airline carriers have colluded in the number of flights that they offer. Using a similar model, Charlie and a co-author, Ying Jiang, study whether these mergers have changed the incentives of large “legacy” carriers to engage in predatory pricing strategies against “low cost” carriers.
In simulating dynamic competition, machine learning and cloud computing tools provided by AWS have allowed Charlie to study airline competition in a richer way than done by many airline studies in economics. For example, Charlie has been able to test many different models of airline competition quickly, rather than relying on just one. Charlie notes, “I can swap out different variables and assumptions to see how things change. This is called specification testing, which plays an important role in the subfield of econometrics called structural econometrics. Under normal circumstances, specification testing can take many months. I can run several iterations in a week—that is new science.”
The Need for Speed
When asked why speed is so important, Charlie replied, “Most of the time, structural econometrics papers take several years to get through review process. So for one thing, cloud computing has the potential to bring science out there more quickly. It might only take me a couple of weeks to develop a revision for the paper. Speeding up the research process speeds up the publication process.
Secondly, both machine learning and cloud computing allow researchers to create more realistic models of competition. “Often,” Charlie says, “economists have to use very simplified models to study dynamic competition, where a small number of state variables are chosen by assumption. Especially in network industries (which include airlines, big box retailers, and cable companies, for example), ad hoc simplification might generate misleading conclusions about the appropriate public policies to apply. The hope is that these newer methods allow us to get closer to reality—our model is super rich and very realistic, which is useful since airline networks are comprised of thousands of routes. Most studies of dynamic airline competition include fewer than twenty state variables. In our predatory pricing study, we include more than seventeen thousand. Speed actually helps us approximate reality in network industries, which are notoriously difficult to study due to their complexity. We can also use all of the data, rather than small subsets. I’ve never seen an economics study that uses the entire Department of Transportation airline data set. That’s what we can do with these methods.”
Charlie is currently a visiting PhD student in UW’s Economics Department. He followed his primary advisor, Yanqin Fan, to the UW when she was hired by the Economics Department in 2013, to continue working on joint projects. Charlie’s use of AWS began when he participated in the UW eScience Institute Incubator program (in the Fall of 2014), and led to further collaboration with AWS.
Support from Amazon & eScience
In order to remain in Seattle and continue his research, he obtained an internship with the Central Economics Team at Amazon, led by Pat Bajari. It was Pat who suggested that Charlie sit in on machine learning classes at the UW. Charlie acknowledges the significant contribution of the UW—the eScience Institute (in particular Bill Howe’s help in connecting Charlie with helpful people) and the UW’s machine learning courses have made to his research “Without machine learning—the UW’s curriculum has to be one of the best in the country—and the eScience Institute (they embraced me as one of their own), this work would not have been possible.”
What’s on the horizon for Charlie’s research? “My dream is to make this data and these methods available to researchers, students, and the DOJ. I would love for people to run their own simulations.”
Learn more about Charlie’s work at his presentation at UW’s Cloud Day, November 12.
UW-IT can help your unit get started with AWS; just send a request to firstname.lastname@example.org.