In this project, I work with a dataset from Zillow that has house prices over the course of several month-years, starting in 1996. The dataset also provides other information such as: zip code, region, county name and city - to name a few. I analyze these series using a time series approach and provide recommendations for an hypothetical real estate firm.
The modeling framework is streamlined as follows: time series analysis, basic and rolling statistics, seasonal decomposition, unit-root test, SARIMAX modeling and prediction. Lastly, we delve deeper into the data, repeating this process with microscopic lenses.
An hypothetical is looking to diversify its portfolio. Thinking of risk-weighted returns, then the firm seeks advisory for spotting the best areas over the course of 12-24 months. The stakeholders of this problem, besides the investment firm, are: general public and outside investors, municipalities and counties and communities in general.
The dataset used is a collection of time series with house prices over the course of Apr/96 through Apr/18. Besides prices, there are the other categorical variables: RegionID, RegionName, City, State, Metro, CountyName and SizeRank.
Splitting and ranking the dataset into the top 10 counties with respect to median prices, we end up with the following locations:
I rank which counties are the most valuable with respect to their median and mean and assessing how they fare with respect to risk (i.e. standard deviation). I also provide the evolution of house prices with respect to price, their 12-month window rolling statistics (mean/std. deviation) and the log returns. From this initial filtering, we end up with 3 counties as the best risk-weighted returns: Suffolk, NY; Palm Beach, FL and Miami-Dade, FL.
I start with testing stationarity and performing a seasonal decomposition to assess the best combination of Autoregression Integrated Moving Average factors (i.e. p, d, q) and Seasonal Factors (p, d, q, s). Given computational constraints, I limit those to 1 round of runs (even though I also provide examples with 2 runs). Then, Autocorrelation and Partial Autocorrelation plots are provided to check the consistency of the seasonal decomposition findings. Lastly, I perform Root-Mean Square tests on both training and testing data - which both provide satisfying results for our three locations.
Using a time series approach I provide a SARIMAX model that is able to perform well both with training and testing data. Root-mean squared errors
are less than 3% of the historical mean prices. Thus, one can infer the model is good to predict with a certain accuracy house prices. The model is
inversely as good as the time horizon increases, so for shorter-term forecast (up to 24-months) we have more precise price estimates.
The best historical risk-weighted return is located in Suffolk, NY, followed by Palm Beach, FL and Miami-Dade, FL. So the investor can pursue two distinct strategies:
- Divide and Conquer - investing in the best bang for the buck county/location all at once;
- Rays of the Sun - choosing one specific location and irradiating the strategy within that county;
- Alternative - investing in more established counties.
I list below some ideas about further research:
- Implementing Portfolio Theory as a way of assessing optimizing investment strategies;
- Adding exogenous features that might explain house prices (such as socio-economic data, school district data, etc.);
- Delving deeper to explain the variation within counties to assess under/overpriced real state.
This repository provides basic guidelines on how to navigate through this notebook and its material.
Here you will find:
- Presentation file;
- Jupyter Notebook;
- README.md file;
- Image Folder;
- Data (.CSV) folder.