Project Goals

• Can we predict neighborhood stability?
• Let's use time since the last sale of a house as a proxy.
• Can we predict this from available property tax data?
• First, can we accumulate enough property tax data?

Web Scraping

• The Will County website has the data we need.
• Extracting data is straightforward if you have the address or PIN.
• But guessing these can be slow.

Web Scraping

Linear Regression

• Can we predict "Longevity" (time since last sale)?
• Not very well.
• The mean "Longevity" is about 13 years.
• The typical error of the better models was 5 years.

Learning Curves

• The model converges quickly.
• More data is likely not going to help.

Learning Curves

• With more data, we continue to see huge variance.

Poorly correlated features?

Correlation Matrix Snippet

• Most of our features have poor correlation with our target.
• Feature evaluation confirms Sale Amount is the strongest signal.

Linear Regression - Assumptions

Have we met the key assumptions for Linear Regression?

• Linear relationship
• Multivariate normality
• No or little multicollinearity
• No auto-correlation
• Homoscedasticity

Homoscedasticity

Collinearity

• There were several groupings or clusters of similar things among the features.
• For each of these all but one of the features were removed.