The regular way to compute the F-value for a Chow forecast test is: $$F=\frac{(e_R'e_R-e_1'e_1)/g}{e_1'e_1/(n-k)}$$
My professor said something today about that a Chow forecast test using $R^2$ would be wrong, because the number of observations in the restricted model that gives $e_R$ is different from that in the model that gives $e_1$. Could anyone please explain me why this is so?
