greg-maggard / zillow_regression_project

Project using regression to improve Zillow's house price prediction model.

Predicting Home Prices With Regression

Greg Maggard
July 22, 2022

=========================================================================================================

General Overview:

This project aims to use regression models to predict home values from the Zillow dataset containing listings from 2017.

Main Goals:

• Use regression machine learning models to make predictions regarding housing prices.
• Identify key features that can be used to create an effective predictive model.
• Use findings to make recommendations and establish a foundation for future work to improve model's performance.

=========================================================================================================

So What and Why?

• Improving this model provides significant opportunity for Zillow to increase their ability to accurately predict the valuation/sale price of a home, which is integral to its ability to attract customer, generate leads, and earn commissions on sales.

=========================================================================================================

Executive Summary:

• This project aims to improve upon an existing model to predict home prices for Zillow listings from 2017.
• Data was acquired by querying 2017 home data from the Zillow database.
• Exploration was done to test some hypotheses about relationships between features.
• Multiple regression models were run, ultimately finding Linear Regression to be the most effect, reducing the model's error by ~19.44%.
• I recommend deploying this model over the original for the time being, given that there is a 20% increase in model performance.
• Given more time, I'd like to continue to refine the model with current features, while also trying to add more home features and data into the dataset.

=========================================================================================================

Project Plan:

Acquisition:

• Querying data from the Zillow database for use in the model.

SELECT 
    bathroomcnt AS bathrooms,
    bedroomcnt AS bedrooms,
    taxvaluedollarcnt AS value,
    calculatedfinishedsquarefeet AS square_feet,
    yearbuilt AS year_built,
    fips,
    latitude,
    longitude,
    lotsizesquarefeet AS lot_size
FROM
    properties_2017
        LEFT JOIN
    predictions_2017 USING (parcelid)
        LEFT JOIN
    propertylandusetype USING (propertylandusetypeid)
WHERE
    propertylandusedesc IN ('Single Family Residential' , 'Inferred Single Family Residential')
        AND YEAR(transactiondate) = 2017;

52,441 rows are returned, with the above 9 columns.

Data Dictionary:

Column/Feature	Description
bathrooom	The number of bathrooms in the home.
bedrooms	The number of bedrooms in the home.
value	The tax-assessed value of the home. Not the home's ultimate sale price.
square_feet	The home's square footage.
year_built	The year the home was built.
fips	"Federal Information Process System" code, used to identify zip codes in the U.S.
latitude	The latitude of the home.
longitude	The longitude of the home.
lot_size	The square footage of the lot on which the home is built.

=========================================================================================================

Wrangling:

- General Data Format Cleaning:

• Converting moving decimal left 6 places on latitude and longitude.

- Assigning county names to each home, based on the fips code one the record:

• 60370: Los Angeles County
• 60490: Orange County
• 61110: Ventura County

- Homes Dropped:

• Having 0 bedrooms
• Having 0 bathrooms
• Having less than 120 square feet
  • These homes to not meet the CA minimum to classify as a residence.
• Having more than 10,000 square feet
  • These homes make up a small portion of the set and could skew the model.
• Home value over 1.6 million dollars
  • These homes make up a small portion of the set and could skew the model.
• Having more than 100,000 square footage of lot size.
  • These homes make up a small portion of the set and could skew the model.
• Home records containing null values in any column.

- Columns Created:

• Column displaying the ratios of bedrooms to bathrooms.

Notes on Wrangling:

• All of these cleaning steps are carried out in the wrangle script, and leave 94.3% of the data remaining.
• I feel comfortable with the omission of this data, as I want to ensure that I'm not excluding too large a chunk of my total set, but do want to be sure that I'm focusing my model on homes that comprise the bulk of Zillow's business.

=========================================================================================================

Exploration:

Key Questions Answered:

• Is there a significant relationship between square footage and home value?
  • Statistical Finding: There is sufficient evidence to reject the null and assert that there is a significant relationship between square footage and the assessed value.
• Is there a significant relationship between lot_size and home value?
  • Findings: There is sufficient evidence to reject the null and assert that there is a significant relationship between lot size and the assessed value.
• Is there a relationship between the county in which a home is built and its value?
  • Statistical Finding: There is sufficient evidence to reject the null and assert that there is a meaningful relationship between home values and county.
• Is there a relationship between the year a home is built and the square footage of a home?
  • Statistical Finding: There is sufficient evidence to reject the null and assert that there is a meaningful relationship between home values and county.

=========================================================================================================

Modeling:

Baseline Model:

• Created a baseline model that uses no features, and simply takes the mean home value as the estimate.
• It's important to note here that there is no "machine learning" happening with this model; all it is doing is finding the mean assessed home value and saving that to a new column in the DataFrame.

Regression Model:

• Created and OLS linear regression model with 9 features to attempt to predict house values.

Modeling Takeaways:

• The OLS Regression model beats the baseline by ~19.81% on the test set.

=========================================================================================================

Conclusion:

Key Findings:

• Home square footage, lot size, county, and build year all proved to be significant drivers of home value.
• The OLS Linear Regression model, with the features provided, was able to improve upon the baseline model by ~19.81%. I would expect that this will be the case on further out-of-sample data.

Recommendations:

• I recommend deploying this linear regression model for now, as an improvement of nearly %20 over the current model means there is substantial financial benefit on the line.
• I'd also recommend perhaps doing more qualitative research to understand what factors customers look for in estimating a home's value.

Next Steps:

• With the luxury of more time, I would like to further explore the variables in the dataset and see if I could find a better combination to refine the model.
• It would likely be worthwhile to look into acquiring more data on the homes to see if there are other factors that could be drivers of home value.
  • There are factors like how recently a home has been renovated, proximity to quality schools or greenspaces, or myriad other aspects that could be considered.

=========================================================================================================

Steps to Reproduce:

• Ensure that you have an env.py file that includes relevant database credentials to query the data.
• Download wrangle.py, evaluate.py, and explore.py files.
• Download and run the zillow_final_report.ipynb file.

=========================================================================================================