UC Berkeley

The thesis consists of two chapters on real estate economics.

In the first chapter, I study the impact of intermediaries in the real estate transactions. In many markets, intermediaries collect a substantial amount of commission in exchange for their expertise. Real estate is a prominent example—Americans paid more than $60 billion for real estate brokerage services in 2014. In this paper, I find a significant positive relationship between listing agents with greater recent experience and sales price, which is entirely driven by the sorting of more experienced agents into better houses. Once both observed and unobserved house characteristics are controlled for, there is no significant effect of experienced listing agents on average sales price. However, I present a novel finding that there is a significant negative relationship between recent listing agent experience and the variance of the sales price, which means that experienced listing agents add value to risk- averse sellers. Moreover, I investigate the mechanism that drives this negative relationship. Looking at individual performance, I show that a listing agent’s past performance predicts his future performance, and that good past performance leads to more listings in the future. The relationship is driven by the survivorship of better agents over time rather than the accumulated expertise of the agents. The channel is consistent with survivorship bias and decreasing returns to the number of listings for listing agents, similar to Berk and Green (2004)[8].

In the second chapter, I geocode a rich real estate repeated sales dataset and map each property to its school district and neighborhood. I study how big data algorithms differ from OLS regression in predictive power and how robust those algorithms are to data stratification. I find that it is computationally expensive for the random forest algorithm to use step functions to approach the linear data generating process. Once there are fewer predictors, the RF algorithm outperforms other algorithms. This is robust to different model specifications. In addition, the random forest algorithm provides similar results under different stratifications. I also study the effect of keywords on sales price and how informative they are in predicting sales price. I find that certain keywords can be valuable in explaining variation in the data but have insignificant impact on the average sales price, suggesting that the interaction between such keywords and other house features together should be considered when we specify our models. Lastly, I am able to exploit cross time variation in school academic performance index to identify the effect of school quality on house prices controlling for neighborhood fixed effect. I find school quality has a robust significantly positive effect on property sales price.