When should I use polynomial regression to capture curvature in the data?

Question

AcadiFi · Accepted Answer

Polynomial regression (adding X², X³, etc.) is useful when you have clear evidence of nonlinearity and a theoretical reason for the curvature. The classic example: diminishing returns on advertising. Sales rise with spend but at a decreasing rate — adding a negative X² captures that concavity.

Craftwood Brewery fits Sales = 42 + 18 × Spend − 0.35 × Spend² and finds R² jumps from 0.68 (linear only) to 0.91, with the Spend² t-stat highly significant. The quadratic captures that doubling ad spend from $50K to $100K adds less revenue than doubling from $10K to $20K.

When polynomials help:
- Clear theoretical U-shape or inverted-U (diminishing returns, Laffer curves, optimal portfolio concentration)
- The data show pronounced curvature in residual plots from a linear fit
- Domain knowledge predicts the specific curvature direction

When to be careful:
- Extrapolation beyond the data range is dangerous — high-order polynomials explode
- Cubic and higher terms often overfit; stop at quadratic unless you have strong justification
- Polynomial terms are highly correlated with the base variable; always center X first to reduce multicollinearity
- A spline or kernel-smoothed curve may fit better than a polynomial for irregular shapes

An alternative to polynomials: log-transforms. If you suspect diminishing returns, try Sales = a + b × log(Spend). This often fits better than quadratic and extrapolates more safely. Compare models via adjusted R², AIC, and out-of-sample RMSE, not just in-sample fit.

When should I use polynomial regression to capture curvature in the data?

Master Level II with our CFA Course

Related Questions

Practice Questions