Trying to understand this relationship I came across this conversation in an email group:

I got a problem while analyzing my data using NN. I got R-square with 0.45 with MSE around 5000. SO I am wonder why? There must be a relation between R-square and MSE?

R2 = 1 - SSE/SS0

SSE = N*MSE SS0 = (N-1)*VAR(Target)

VAR(T) = SS0/(N-1) = [SSE/(1-R2)]/(N-1)

= [N/N-1]*MSE/(1-R2)

~ 5000/0.55 ~ 9091

If you had standardized your targets to unit variance you would have obtained MSE ~ 0.55

Is that correct? What insights can we reach understanding this relationship?


1 Answer 1


Yes, allow me to elaborate.

Recall that for some outcome $y_i \in \mathbb{R}, \forall i=1,2,..,n$ we define MSE and $\textrm{R}^2$ as

\begin{equation} \textrm{MSE}(y, \hat{y} ) = \frac{1}{n} \sum_{i=1}^{n} (y_i - \hat{y_i})^{2} \end{equation}

\begin{equation} \textrm{R}^2(y, \hat{y} ) = 1 - \frac{\sum_{i=1}^{n} (y_i - \hat{y_i})^{2}}{ \sum_{i=1}^{n} (y_i - \bar{y})^{2} } \end{equation}

So, as you noted, $\textrm{R}^2$ is a normalized version of MSE, we use MSE for reporting because I think it's a simple metric and it is technically the loss-function we are minimizing when we solve the normal equations.

$\textrm{R}^2$ is useful because it is often easier to interpret since it doesn't depend on the scale of the data.

As a concrete example, consider two models: one predicting income and the other predicting age, $\textrm{R}^2$ will make it easier to state which model is performing better.*

*In general, this isn't a great idea and you shouldn't compare metrics like $\textrm{R}^2$ across different models to make these sorts of claims because some things are just fundamentally harder to predict than others (e.g., stock markets vs. who survived the Titanic).


Not the answer you're looking for? Browse other questions tagged or ask your own question.