RMSE Expresses Typical Error in Response Units

Root mean squared error summarizes residual size:

$\mathrm{RMSE}=\sqrt{\frac{1}{n}\sum_{i=1}^n e_i^2}.$

Because the square root reverses the squared scale, RMSE has the same units as the response.

Residual standard error

For an estimated regression with $p$ coefficients,

$s=\sqrt{\frac{\mathrm{SSE}}{n-p}}$

uses residual degrees of freedom. Under standard assumptions, $s^2$ estimates the error variance.

Context gives meaning

An RMSE of 2 is excellent if acceptable error is 10 units and poor if acceptable error is 0.1 units. Compare it with measurement resolution, natural variation, and the cost of mistakes.

Calculate it from residuals

For residuals $-2,1,3$, the squared errors are $4,1,9$. Their mean is $14/3$, so

$\mathrm{RMSE}=\sqrt{14/3}\approx2.16.$

The mean absolute error is $(2+1+3)/3=2$. RMSE is larger because the residual of magnitude $3$ receives disproportionate weight.

Sampling denominator versus prediction score

Use $n-p$ when estimating the underlying noise variance after fitting $p$ coefficients. Use the number of held-out predictions when reporting predictive RMSE. These answer different questions and should be labeled clearly.

Check your understanding

Can RMSE be zero while the model is scientifically wrong?

Show the reasoning

Yes. An overly flexible model can interpolate the training data exactly, or leaked information can make evaluation invalid. Zero training error does not establish generalization or scientific validity.