I'm trying to understand fraction of unexplained variance on wikipedia: https://en.wikipedia.org/wiki/Fraction_of_variance_unexplained

$$\text{unexplained variance} (y, \hat y) =\frac{\operatorname{Var}\{y-\hat y\}}{\operatorname{Var}\{y\}}$$

Correct me if I'm wrong, it's basically find out the variance between the difference of the expected value $y$ and predicted value $\hat y$, where $y$ is the real value, and $\hat y$ is the predicted value.

In a linear regression setup: $\hat y=mx+c+e$, where $e$ is the error term (residual or unexplained error). So if $e$ is $0$ for all predictions, then the unexplained variance would be $0$, since there is no variance.

However, $\operatorname{unexplained variance} (y, \hat y) =\dfrac{\operatorname{Var}\{y-\hat y\}}{\operatorname{Var}\{y\}}= \frac{SS_\text{err}}{SS_\text{tot}}$ is the confusion part.

$$SS_\text{err}(\text{Residual Sum of Squares})=\sum_{i=0}^n (y_i-\hat y_i)^2$$

$SS_\text{tot}(\text{Total Sum Of Squares})=\sum_{i=0}^n (y_i-\bar y_i)^2$, where $\bar y_i$ is the average value.

$\operatorname{Var}\{y-\hat y\}=E[((y-\hat y)-\sum_{i=0}^n(y-\hat y))^2]$, somehow this does not look like Residual sum of squares?


1 Answer 1


You did most of the work yourself. From the final line you wrote:

$\operatorname{Var}\{y-\hat y\}=E[((y-\hat y)-\sum_{i=0}^n(y-\hat y))^2]$

This isn't quite correct, remember that the variance of some $X$ is the expected value of $X^2$ minus the expected value of $X$ squared, i.e. $var[X]=E[X^2]-E[X]^2$

Therefore the variance of $y-\hat y$ is:

$var[y-\hat y]= E[(y-\hat y)^2]-E[y-\hat y]^2$

$var[y-\hat y]= \frac{1}{n} \sum_{i=1}^n(y_i-\hat y_i)^2- \left(\frac{1}{n}\sum_{i=1}^n(y_i-\hat y_i)\right)^2$

In linear regression we always set the parameters so that they overestimate some $\hat{y}$ and underestimate some $\hat{y}$, the amount of overestimation and underestimation is equal. Therefore, when you sum over all $y-\hat{y}$ the total is $0$, this simplifies things a lot. Although I mentioned linear regression, most other forms of regression do this too.

$var[y-\hat y]= \frac{1}{n} \sum_{i=1}^n(y_i-\hat y_i)^2$

Clearly $var[y-\hat{y}]=\frac{1}{n}SS_{err}$

Likewise, $var[y]=\frac{1}{n}SS_{tot}$

Hence $\frac{var[y-\hat{y}]}{var[y]}= \frac{SS_{err}}{SS_{tot}}$

  • $\begingroup$ Thanks, did not know the total is 0. I was stuck for a day already. $\endgroup$ Commented Mar 6, 2017 at 16:57

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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