Oral Defense of Doctoral Dissertation – Computational Social Science – Rethinking Housing with Agent-based Models: Models of the Housing Bubble and Crash in the Washington, D.C. Area 1997-2009 – Jonathan Goldstein

Notice and Invitation
Oral Defense of Doctoral Dissertation
Computational Social Science
Department of Computational and Data Sciences
College of Science
George Mason University

JONATHAN GOLDSTEIN
Bachelor of Arts, Cornell University, 2004

Rethinking Housing with Agent-based Models: Models of the Housing Bubble and Crash in the Washington, D.C. Area 1997-2009

Monday, April 24, 2017, 10:00-Noon
Research Hall, Room 92

All are invited to attend.

Committee
Robert Axtell, Chair
William G. Kennedy
Carlos Ramirez
J. Doyne Farmer

This dissertation presents a series of related agent-based models (ABMs) of the housing market of the Washington DC Metropolitan Statistical Area. The models investigate the causes of the housing market bubble and crash during the time period 1997-2009 and policies that could have avoided such a crisis. The work in this dissertation contributed to three research areas: understanding the underlying causes of the housing crisis, demonstrating the utility of ABMs for investigating macro phenomena, and improving ABM methodology.

Using the housing market models, I ran counterfactual analyses to investigate causes and potential mitigations of the crisis. I show that leverage and expectations are the two most prominent causes of the bubble, but that other factors, such as interest rates, the norm governing the share of income to spend on housing, and seller behavior influence the bubble. I find that lending standards and refinance rules play almost no part in the bubble, contrary to some theories of the housing crisis. Towards the end of the dissertation, I pair the housing market with a model of mortgage-backed securities. I show that the increased velocity of lending made possible by securitization can increase the size of bubbles and make markets more fragile, increasing the likelihood of crashes.

The ABMs in this dissertation exploit multiple large, heterogeneous data sets and utilize realistic behavioral rules to match detailed housing market dynamics. Input data include loan level data, multiple listing service records, and demographic information from a variety of sources. The ABMs exploit this data by choosing the precise areas of input distributions to use based on the context of the model. This allows the ABMs to match not only aggregate outputs, but intermediate outputs and data distributions. For example, the ABMs in this dissertation not only reproduce empirical macro phenomena, such as the shape of the house price index, but also intermediate variables (e.g., distribution of loan types, average leverage, average days on market, average ratio of sold price to original listing price) and output distributions (e.g., distribution of house prices).

Throughout the dissertation, I follow several methodological principles in construction and analysis of the ABMs. First, I demonstrate the use of data to constrain the models as much as possible. Next, I describe a sensitivity analysis methodology that goes beyond parametric variations, but also varies model rules in what I term a structural sensitivity analysis. I demonstrate how criticisms about ABMs with regard to their opacity, brittleness, and dependency on arbitrary modeling decisions can be resolved through such an analysis. I also describe the architectural design of the models, which makes explicit the theoretically-inspired behavioral rules, facilitating structural sensitivity analyses.

A copy of Jon’s dissertation is available for examination from Karen Underwood, Department of Computational and Data Sciences, 373 Research Hall. The dissertation is available to read only within the Department and cannot be taken out of the Department or copied.