The goal of this project was to create a model that would build 95% confidence intervals for housing prices in King County, WA. Our dataset is modified from the original available on kaggle.
The columns that we used to build our regression model were:
Response:
- price - The sale price of the house (in dollars)
Predictors:
- sqft_living: The amount of living space, in square feet.
- sqft_lot: Used in conjunction with sqft_living to create a Floor Area Ratio.
- sqft_lot15: The square footage of the land lots of the nearest 15 neighbors
- waterfront: Indicates whether the house is on the waterfront (1) or not (0).
- bathrooms: The number of bathrooms, including partial bathrooms that do not have specific fixtures (i.e. a bathroom with a sink and toilet only is .5)
- bedrooms: The number of bedrooms
- floors: The number of floors, where 1/2 indicates a partially-finished floor or a loft
- zipcode: The house's zipcode
We also used the MapQuest API to retrieve city information, based on the columns lat and long. [code]
This repository contains the following files:
- MappingFunctions.ipynb - Contains lat/long scatterplots using hues to display patterns in various predictors.
- EDA.ipynb - Contains visualizations and numerical summaries from our initial exploratory data analysis.
- Regression.ipynb - Contains functions for transforming the data, including taking the log and one-hotting, in addition to performing the actual regression analysis.
- predicting-housing-prices.pdf - The slides from our final presentation on July 7, 2020.
The data folder contains the following files:
- kc_house_data_original.csv - The original, unmodified dataset.
- updated_housing_data.csv - A modified version of the original csv which contains new columns for ratios and city information.
Through our analysis, we were able to construct 95% confidence intervals for housing prices. Approximately 94% of the data was for houses that sold for less than $1,000,000 (in 2014-2015), and approximately 99% of the data was for houses that sold for less than $2,000,000, so the model's accuracy decreases for more expensive houses due to lack of available training data.
We built the model by splitting the dataset into 80% training and 20% testing. After a lot of tinkering and trial and error, we got residuals that we were happy with:
Once we knew we were looking at a solid set of columns as predictors, we retrained the model on the entire dataset and used it to predict prices of houses on Zillow.
Once the regression model was finished, we used the same columns within a sequential neural netowrk architecture and compared the results.
