What is the fastest way if I am only interested in computing overlapping regression estimates recursively? I have daily stock return data as follows. The following code automatically downloads and generates the MWE data.
data adjusted(drop=dummy);
input ticker $ @@;
url=compress("https://query1.finance.yahoo.com/v7/finance/download/"||
ticker||"?period1="||dhms("1jan1901"d,0,0,0)-315619200||
'&period2='||dhms("31dec2100"d,23,59,59)-315619200);
infile dummy url filevar=url firstobs=2 dsd truncover end=last;
do until(last);
input @;
if index(_infile_,"null") then input;
else do;
input date yymmdd10. +1 (4*dummy)(:$1.) adjusted dummy;
output;
end;
end;
cards;
^gspc ^irx ba dis hpq ibm ko
;
data return(drop=adjusted);
set adjusted;
if ticker="^irx" then return=lag(adjusted/100)*(date-lag(date))
/(intnx("year",lag(date),1,"same")-lag(date));
else return=ifn(ticker=lag(ticker),adjusted/lag(adjusted)-1,.);
run;
proc sql;
create table rirf as
select i.ticker,i.date,i.return-j.return as rirf,
k.return-j.return as rmrf
from return(where=(ticker not in ("^gspc","^irx"))) i
join return(where=(ticker="^irx")) j on i.date=j.date
join return(where=(ticker="^gspc")) k on i.date=k.date
order by ticker,date;
quit;
which look like
Obs ticker date rirf rmrf
1 ba 19620102 . -0.85%
2 ba 19620103 1.99% 0.23%
3 ba 19620104 -0.99% -0.70%
4 ba 19620105 -1.99% -1.39%
5 ba 19620108 0.23% -0.80%
6 ba 19620109 0.24% 0.04%
7 ba 19620110 -0.01% -0.28%
8 ba 19620111 1.00% 0.59%
9 ba 19620112 2.73% 0.34%
10 ba 19620115 -1.96% -0.22%
and for each ticker in each month, I want to regress rirf on rmrf using the trailing 12-month data. In December 1962, I want to compute reg estimates (I only need the beta) for each ticker using the data from January 1, 1962 to December 31, 1962. The estimates in January 1963 will be based on the data from February 1, 1962 to January 31, 1963, and so forth. Using macro as follows will be easy but inefficient.
%macro reg;
%let date=%sysfunc(inputn(31dec1962,date9));
%do %until(&date>%sysfunc(inputn(30sep2019,date9)));
proc reg data=rirf(where=(%sysfunc(intnx(month,&date,-11))<=date<=&date))
noprint outest=reg(keep=ticker rmrf);
by ticker;
model rirf=rmrf;
quit;
%if &date=%sysfunc(inputn(31dec1962,date9)) %then %do;
data beta;
set reg;
date=&date;
run;
%end;
%else %do;
data beta;
set beta reg(in=reg);
if reg then date=&date;
run;
%end;
%let date=%sysfunc(intnx(month,&date,1,end));
%end;
%mend;
%reg
I can instead tabulate the ingredient data first for reg and then feed the data with by statement. This is much faster than the first one, but the intermediate rirf2 table will require gigantic disk space.
data rirf2;
date="31dec1962"d;
do until(date>"30sep2019"d);
output;
date=intnx("month",date,1,"end");
end;
run;
proc sql undo_policy=none;
create table rirf2 as
select i.date,ticker,rirf,rmrf
from rirf2 i join rirf j on j.date<=i.date<=intnx("month",j.date,11,"end")
order by date,ticker;
quit;
proc reg noprint outest=beta2(keep=date ticker rmrf);
by date ticker;
model rirf=rmrf;
quit;
I can also think of an IML version that repeats use and close over again doing inv(x`*x)*x`*y recursively. There will be tones of different ways, but what is the fastest approach unless I assume parallel computing? I think movstd in proc expand seems efficient, but, as far as I know, proc expand only deals with one variable. I used the first macro approach in the past, am using the second SQL approach now, and want to switch if there is anything even faster in SAS. I wonder if there is a better version from more experienced experts.
