Does a quadratic model use two predictors?

It uses two polynomial columns, x and x squared, derived from one original predictor, plus an intercept in the usual form.

Should the highest degree always be selected?

No. Additional flexibility can overfit and worsen predictions.

Does centering change the curve?

With an equivalent complete polynomial basis and correctly transformed coefficients, it changes the representation but not the fitted curve.

AI Glossary / Statistics and probability / Intermediate

Polynomial regression

Definition

Polynomial regression models an outcome using powers of one or more numerical predictors. A quadratic model with one predictor has the form E[Y|X=x] = beta_0 + beta_1 x + beta_2 x^2; higher-degree models add further powers. Although the fitted curve can be nonlinear in x, the usual model remains linear in its unknown coefficients. It can therefore be fitted with linear-regression methods when their assumptions and the design are appropriate. Polynomial terms offer a way to represent curvature over a relevant range.

What it means

A straight line assumes a constant change in the mean outcome for a unit change in the predictor. A polynomial allows that change to vary. A quadratic can bend upward or downward, while higher powers allow more complex shapes. Flexibility is not automatically a benefit: fitting a complicated curve to limited data can reproduce noise and behave poorly on new observations. The degree and predictor range are part of the model specification, and predictions outside the observed range can be especially unstable.

How it works

Choose a plausible polynomial degree and construct the predictor terms, often after centering or scaling x. Fit the coefficients using an appropriate loss or statistical model. In ordinary least squares, this means minimizing squared residuals across the intercept and polynomial columns. Examine residual patterns, influential observations and predictive performance. Retain lower-order terms when using the usual hierarchical specification, unless a justified constraint defines a different model. Select complexity without using the final evaluation data. For multiple predictors, distinguish separate powers from interaction terms, since they represent different aspects of the response surface.

Three practical examples

Process modeling: approximate how average output changes over a studied range of operating temperature.
Demand analysis: describe a curved association between a numerical feature and observed demand without assuming that the relationship is causal.
AI baselines: add polynomial features to a regression model and compare its held-out prediction with a straight-line baseline and other flexible methods.

Think of it this way

Think of choosing a ruler that can bend in a controlled way. A slight bend can capture a pattern that a straight ruler misses. Allowing many bends can trace every irregularity in a small sample, but those irregularities may not recur. The useful curve is the one that represents the task reliably, not simply the one that passes closest to every training point.

Worked example

Suppose an original artificial fitted mean curve is y_hat = 10 + 4x - x^2 over a studied range from 0 to 4. Predictions at x = 0, 1, 2, 3 and 4 are 10, 13, 14, 13 and 10. A one-unit increase from 1 to 2 changes the prediction by +1, while the increase from 2 to 3 changes it by -1. The coefficient 4 is therefore not a constant four-unit effect across the range. The derivative is 4 - 2x, and the curve's maximum occurs at x = 2. If z = x - 1, exactly the same curve becomes 13 + 2z - z^2: centering changes coefficients but not fitted predictions. Extrapolating to x = 6 gives -2, which could be impossible for a nonnegative outcome. This illustrates why a sensible fit within the observed domain need not justify distant predictions.

Illustrative case study

In a fictional case, Brook Manufacturing sees curvature in an output-versus-temperature plot. Analysts compare linear and quadratic fits using a planned validation procedure and examine residuals. The quadratic improves prediction within the studied operating range, so they document that range and retain the linear term along with the square. When management asks for predictions at a much higher temperature, the team requests relevant experimental evidence rather than treating the curve's extrapolated turning points as proven process behavior.

Common misconceptions and mistakes

Polynomial regression is not necessarily nonlinear regression in the parameter-estimation sense; the ordinary coefficients enter linearly. A high training R-squared does not establish generalization. The coefficient of x is a local slope at x = 0 in a quadratic parameterization, not a constant effect when higher powers are present. Centering can improve numerical conditioning and interpretation, but it does not create information or automatically solve every collinearity problem. Orthogonal polynomial coefficients depend on the chosen basis and should not be read as raw-power coefficients. A polynomial association also does not establish a causal response to changing x.

Business use

For businesses, a low-degree polynomial can provide an understandable baseline or a local approximation for planning. Specify the valid operating domain and report prediction uncertainty where it matters. If the outcome is constrained, check whether the model produces impossible values. Compare with alternatives such as splines when local flexibility is needed, and use validation that respects time or grouping in the data. Store the complete transformation and coefficients at sufficient precision so deployment reproduces the fitted curve. A chart trendline with heavily rounded coefficients may not reproduce the original predictions accurately.

Check your understanding

In y_hat = 10 + 4x - x^2, is the predicted change from x = 2 to x = 3 equal to 4?

Answer: No. The prediction moves from 14 to 13, a change of -1, because the squared term changes too.

Related terms

Linear regression
Overfitting
polynomial features
quadratic model
interaction
extrapolation
spline

Learn next

Learn how derivatives describe local slopes, then compare raw powers, centered predictors and orthogonal polynomial bases before studying flexible alternatives such as regression splines.

Sources

Penn State, STAT 462, Polynomial Regression: https://online.stat.psu.edu/stat462/node/158/ Penn State, STAT 501, Data Transformations and Polynomial Regression: https://online.stat.psu.edu/stat501/Lesson09 Scikit-learn, Polynomial regression: extending linear models with basis functions: https://scikit-learn.org/stable/modules/linear_model.html#polynomial-regression-extending-linear-models-with-basis-functions

Previous: Polynomial matrix spectral factorization Next: Polynomial trend coefficient