well, try something simple like this for the cumulation:
> data returns;
> 	set crsp.msf;	/* this is our master dataset of returns/prices */
> 	logret = log(1+ret);
> 	keep cusip date logret;
> 
> proc sort;
> 	by cusip date;
> 
> proc means;
> var ret;       [I think he meant logret here  -- doncram]
> by cusip;
> output out=final sum=lcumret;
> 
The output dataset, final, now has summed log returns by cusip in 
variable LCUMRET [ie, final looks like CUSIP _TYPE_ _FREQ_ LCUMRET]
> 
> data cumulate;
> set final;
> cumret=exp(lcumret);
> keep cusip cumret; /* this drops lcumret, and the _type_ and _freq_ 
> 			created by proc means */
> 
> proc print;
 
Now, suppose you wanted to do the same thing by years.  Define a YR 
variable, and do the proc sort with a BY CUSIP YEAR and means with a 
CLASS CUSIP YR --  The only difference will be you'll get extra 
observations in the output dataset with _type_ set to different 
values-- discard the irrelevant ones.