BTW, I see that Stata is very restrictive in this department. With "predict" after "xtreg" you can't get an out of sample forecast even in the time dimension when no time dummies are present -- other than plain "xb" which omits individual effects, whether fixed or random.