Random Forest Regression to Impute Missing Values

Trying to impute missing Gross Square Footage for some NYC buildings, I first tried the best regression model with interactions of year, age, building class, tax class, number of stories, square footage before apportionment … I tried the best model chosen by BIC, Elastica Net, and Lasso. The best pseudo out of sample R-squared (not the best measure here, but we all understand it) was 40 percent.

There are so many other interactions that I hadn’t included for model selection, so Random Forest Regression outperforms my linear regression; surprise, surprise! With a 78 percent R-squared in my test set I was ready to roll out imputing missing values. Mean percentage absolute error (MPAE) was 20 percent which was driven by some outliers.

I’ll add some fun details here first chance.