Variable creation in time series regression I have a coworker who wants to create his own independent variable to add to a time series regression model because he believes that his variable will encapsulate more information. Is this advisable? His variable is essentially a weighted average, but I have concerns about the variable's interpretation (it is kind of odd), its behavior, its distribution etc. Any thoughts on this? The math major part of me felt slightly uncomfortable with this idea!
 A: There is no problem with creating new independent variables in a model. This is often the hardest and most crucial part of building a model. You can simply run your model with and without this new variable and compare the two outputs using some statistical tests. See how the performance improves/worsens and then you can make an informed decision on if the new variable is helpful or not. 
A: I agree with mike1886. I would also add that there is often the tendency for these new variables to be influenced to some extent by one's reading of patterns in the time-series, in which case they are not longer truly independent variables.
As much as possible try to base the development of these on independent data, or at least make them up before looking at the data.
For example, a marketing department might feel that blips they see in sales patterns must have been related to marketing campaigns that they remember doing, but using monthly advertising budget, or numbers of ads run per month as a surrogate will afford you a fair amount more confidence in your conclusions.
