This weekend I was exploring the competitions section of the Kaggle.com website. One competition that interested me was the competition to predict house prices based on a dataset from Ames, Iowa that includes 79 explanatory variables. The dataset page gives a brief explanation of these 79 explanatory variables but you have to register to get some more detail on these variables, which consists mostly of the values that each variable can take on. Anyone involved in residential real estate might find the list of potential explanatory variables quite interesting to consider as factors that might affect the price of a house (i.e., housing price indicators).

One aspect of the dataset that I find interesting is that all the variables can be considered "intrinsic" to the house that is being predicted. There aren't any socioeconomic variables like net migration, economic growth or wage levels included. Perhaps this makes sense when we consider that the dataset is only for Ames, Iowa so the same socioeconomic variables arguably apply to all the houses in that area. The validation of the predictive models is limited to predicting Ames house prices using a subset of test data that is "held back" from the training data. A predictive model is good to the extent that it can capture most of the variance in the test data based on the model learned from the training data. There is no further validation that says the same predictive model also works in Houston, Orlando or Toronto.

In a previous blog called Residential Housing Valuation I proposed adding a causal model to our real estate valuation model so that we might incorporate some of these socioeconomic factors that determine housing prices. If your interest in residential valuation is confined to one neighborhood then including these socioeconomic factors might not be that useful as all houses are subject to the same factors; however, if you want to predict residential house prices across diverse neighborhoods then including socioeconomic factors is probably required.

I am impressed with the vast array of mathematical techniques that competitors applied to the problem of predicting house prices. These techniques appear to be based on the assumption that house prices are determined by features intrinsic to the house; however, when I step back I have to wonder whether the predictive models that work for Ames, Iowa is going to apply to an area that is experiencing economic growth, in-migration, and higher wage levels. Will we need to incorporate socioeconomic factors into a more general model for predicting house prices?

We also select realtors based on the perception that they will get as a better price for our property than another realtor and perhaps that is another factor that determines house prices. Does it? If so, how much of an effect might it have?

This blog is an attempt to make sense of what a predictive model of house prices might include. I think the generality of the models in this competition is questionable because they don't include socioeconomic factors that are likely to moderate housing prices from one neighborhood to another. I can also think of non-intrinsic attributes like the realtor used, the listing method used, and the hotness of the market that might have a bearing on predicting house prices at closing. Perhaps they don't matter, but my future research will be looking into seeing if they do matter and to what extent.

Update: Today I came across a relevant study on the accuracy of Zillow's housing price predictions (which they call a zestimate). Zillow does not profess to be that accurate (within 20% of the sale price) and in fact is even less accurate in certain markets like New York. A local realtor claims this is because the market is "hot" but what constellation of features leads to a real estate market being considered "hot"? Interest rates and government policies can "cool" these markets so should they also be included in the definition of a "hot" real estate market?