Uncovering spatio-temporal patterns in Los Angeles house prices: New data, new methods, new findings

House prices are of fundamental importance, yet remain little understood. The importance of this gap in knowledge has been dramatically illustrated by recent events, such as the sub-prime crisis, and the use by Congress of tax-payer guarantees on mortgages that remain at risk, and whose value depends on house prices.;The goal of the thesis is to move forward scientific understanding of house prices. The critical challenge lies in the massive data involved in understanding house price dynamics. There have been very few studies published using anything like the full volume of collected data, and the analysis that have been conducted have largely been rudimentary, given that economists have less experience with these massive data sets than do computer scientists. My Ph.D thesis represents the most complete effort to join intellectual resources with computer scientists that has yet been undertaken in the literature on house prices, and possibly in the broader economic literature.;In chapter 1, I summarize a unique and truly massive new data set that I have put together on house prices in Los Angeles. In addition to almost all housing transactions over the past twenty five years, the data set includes important details on the housing units and on local demographics. In chapter 2, I analyze how well standard repeat sales indices are able to track actual transactions prices. The results are striking: the error term in the standard repeat sales L.A. index displays strong patterns with turnover time, initial prices and geography. In chapter 3, I outline a new methodology for estimating a non-parametric price surface which uses information on geography in a far richer manner than do existing methodologies. I show that accounting for geography eliminates almost all patterns in the error term that the standard index displays. Further, the non-parametric price surface shows a dramatic improvement in prediction accuracy, most strikingly in late 2007, when house prices declined significantly. This research highlights the rich vein of new information that will be available to economic model-builders once we have adapted the machine-learning techniques of computer science to the understanding of house prices.