Rolling regression-firm specific volatility and skewness

[MarinaMag]

Hello,

I am trying to compute firm specific volatility and skewness using monthly stock returns calculated from CRSP monthly stock prices. To be included in the return tests for July of year t, a firm must have Compustat/CRSP data for December of year t − 1 and June of year t. It must also have monthly returns for at least 24 of the 36 months preceding July of year t in order to calculate the firms’ volatility and skewness and the firm’s beta. Furthermore, I will proceed with Fama-MacBeth rolling procedure.

I have found one macro for rolling regression (see attachment) and call it as it is below:

%rollingreg (data= Crsp_Comp, out_ds=_outest_ds , id=permno , date=date , model_equation= ex_ret_mon= market_premium ,

start_date=1980 , end_date= 2010, freq=month, s=1, n=36,regprint=yes);

Then I saved the residuals of the regression and compute the standard deviation of the residuals for the volatility measure and skewness respectively but does not seem to work.

Can anyone help me with that?

Thank you in advance.

[mkeintz]

By "returns test for July of year t" do you mean a window ending June 30 of year t, or do you mean a window endow July 31 of year t?

Do you only want the "July" windows, or do you want windows for every month?  And if you mean every month, then are you still saying you want December data for year t-1, or do you merely mean 7 months prior to the window ending month?

It appears you want 36-month long windows, but will accept up to 12 monthly "holes" in that 36 months, correct?  Which means, at the beginning of your time series, you can have a 24-month window, then a 25-month windows, etc.

For a company with complete monthly data for Jan 1980 through Dec 2010, (31 years), there will 31*12 - 23 = 349 regressions for each PERMNO.  So if you have, say, 500 firms, you'll be constructing 174,500 windows.   PROC SQL, which is in the macro you found, is not an efficient way to prepare this data.  Since crsp/compustat data are almost certainly sorted by permno/date (they are at WRDS), you'll be better off either using a data step make window raw data set view, or possibly a view+proc expand to create rolling sums-of-squares for the regression.