Using rolling regression

[Pan Li]

Dear all,

I run a rolling regression for each equity every 12 months by using a market model.
The purpose of my rolling regression is to obtain the standard deviation of residuals (idiosyncratic risk) and standard deviation of predicted values (systematic risk) for each firm every year.
I know that rmse stands for the standard deviation of residuals.
However, could anyone tell me what code can be used for standard deviation of predicted values.

Thank you!

[Clyde Schechter]

There is nothing returned in e() after regress that gives you the standard deviation of the predicted values. You will have to calculate it directly after the regression:

predict yhat, xb
summ yhat

and the standard deviation will be in r(sd).

Since you are trying to do this on a rolling basis, if you are currently using a -regress- command with the -rolling- prefix, you will probably need to instead write a short program to use with -rolling:-. Or, you could do this using Robert Picard's new -rangerun- command (available from SSC) instead of -rolling:-. Most likely -rangerun- will execute much faster if you are working with a very large data set.

[Pan Li]

Hi Clyde,

Thank you.
But I didn't use codes to execute the regression.
I run the regressions through "statistics">> "Time series" >> "Rolling-window and recursive estimation"
The following is the screenshot.
I don't know how to write expressions for the standard deviation in the red box.

[Clyde Schechter]

OK. I don't think you can do this just using the GUI. You're going to have to use code in a do-file, and you're going to have to write a program. Not having shown any example data, I can only offer limited help, with an example based on a publicly available Stata data set:

clear*

capture program drop myprogram
program define myprogram
    regress mvalue kstock
    gen idosyncratic_risk = e(rmse)
    predict xb, xb
    summ xb
    gen systematic_risk = r(sd)
    drop xb
    exit
end


webuse grunfeld
rangerun myprogram, by(company) interval(year -2 0) use(mvalue kstock)

To do this, you will have to install Robert Picard and Nick Cox's -rangerun- program. Run -ssc rangerun-. I believe that you also need to install -rangestast- in order for -rangerun- to work. (-rangestat- is also from SSC.)

Evidently you will need to modify -myprogram- to customize it to your own situation. Similarly, in the -rangerun- command itself, I have set the window to 3 years (from 2 years before to the current year, inclusive): you are working with monthly data and will want -interval(monthly_date, -11, 0)- or -interval(monthly_date, -12, -1)- depending on whether your 12 month window ends with or just before the current month.