已经想通了，呵呵。。。。

请教下楼主原因，我也遇到了类似的问题。万分感激，虽然帖子问的晚了点儿！

Case by case !

My best guess is that sometimes constant can improve performance and sometimes reduce performance

(1) constant absorbs any kinds of errors in the regression

(2) constant makes sure all residuals have mean of zero

(3.1) If you do not include the constant, the regression is forced to go through the origin such as (0, 0). This means that all of the X and the Y must equal zero at that point. So if your fitted line pass through the origin, the regression coefficients and predictions will not be influenced a lot, and even the overall performance can be enhanced.  

(3.2) If your fitted line does not pass through the origin, the regression coefficients and predictions will be biased if don't include the constant.

This is what I know, just for your reference.