Liquidity in Residential Real Estate Markets with Anthony Lee Zhang

R&R, Review of Financial Studies

We build a rich panel dataset tracking two measures of housing market liquidity: time-on-market and price dispersion. The two measures co-vary closely at seasonal and business-cycle frequencies, but there is substantial independent variation in the cross-section of counties. This suggests that the two measures reflect different dimensions of market liquidity. Using a housing search model, we show that time-on-market and price dispersion can be thought of as equilibrium outcomes from a supply and demand system for liquidity. Consistent with the model’s predictions, proxies for liquidity supply are negatively correlated with both measures, whereas a proxy for liquidity demand is negatively correlated with time-on-market, but positively correlated with price dispersion.

Dynamic Selection in Search Markets: Evidence from the US Housing Market with Zi Yang Kang

This paper studies how the quality of goods traded in a market evolves in the presence of predictable cyclical changes in market conditions. We document a set of novel stylized facts about the US housing market. We show that housing markets of better neighborhoods are more seasonal and higher quality houses are more likely to be traded in summer when housing markets are hot. We then consider a choice problem of a seller who owns an asset with perfectly observable quality who must decide whether to enter the market now or wait for the market conditions to improve. Market participation is costly, and the cost depends on the seller's type. In the model, the quality of sellers who enter the market fluctuates in response to cyclical changes in market conditions. We characterize conditions under which each seller type strategically chooses to delay entering the market until market conditions improve.

A Deep Learning Approach to Predicting Race Using Personal Name and Location (GitHub)

I train a recurrent neural network to predict individuals’ race using information contained in their name and location of residence. I introduce a novel data source that contains millions of race/name/zipcode triplets and covers the entire US. I train my baseline LSTM model using only personal name data. The baseline model attains overall accuracy of 0.87, which is a non-trivial improvement over the existing benchmarks in the literature. However, because personal names of Non-Hispanic Whites and Non-Hispanic Blacks follow similar naming conventions, the baseline model struggles at accurately classifying Non-Hispanic Blacks. Using location data in addition to personal name improves classification accuracy of the model for Non-Hispanic Blacks substantially, and allows the best model to achieve 0.89 accuracy on the test set.