Work / School / 2026
MARS Housing Price Model
STA 4241 Statistical Learning · University of Florida · final project
Compared four regression models on 2,930 Ames, Iowa home sales. MARS with interactions cut test RMSE from $35,976 to $26,737 and still explained itself in plain English.
The question
Home prices don’t move in straight lines. A point of build quality is worth more at the top of the scale than at the bottom. Can a model find those bends on its own without turning into a black box?
Approach
Benchmarked OLS, degree-2 polynomial regression, additive MARS and MARS with two-way interactions on a held-out test set. MARS places “hinge” knots where the slope changes, then prunes itself back with generalized cross-validation.
What I found
Each step up in flexibility helped. MARS with interactions landed at $26,737 RMSE, 25.7% better than OLS. The knots read like market insight: above an Overall Quality of 7, each extra point adds roughly $43K, versus about $9K below it. Homes built after 2007 behave like a different market entirely.
Why it matters
Accuracy you can explain to a non-technical stakeholder. A realtor or lender can act on “quality above 7 is where value jumps.” They can’t act on a coefficient table.