One can %include SAS code (1) without downloading as follows. /*a little delay*/ filename i url "https://raw.githubusercontent.com/junyongkim/sas/master/sganno/usrecm.sas"; %include i; Or (2) the same code after downloading as follows. /*no delay*/ filename j temp; proc http url="https://raw.githubusercontent.com/junyongkim/sas/master/sganno/usrecm.sas" out=j; run; %include j; I found that (1), unlike (2), experiences a little but clearly noticeable delay regardless of the Internet connection speed? Why is (1) a bit slower than (2)? ... View more

I use historical prices to compute returns and cumulative returns as follows. data ko(drop=dummy); infile 'https://query1.finance.yahoo.com/v7/finance/download/ko? period1=-9999999999&period2=9999999999' url firstobs=2 dsd truncover; input date yymmdd10. +1 (4*dummy)(:$1.) ko; return=ko/lag(ko); run; proc expand method=none out=ko; id date; convert return=cumulative/tout=(cuprod); run; And I wanted to shorten the code using retain as follows but failed. data ko(drop=dummy); infile 'https://query1.finance.yahoo.com/v7/finance/download/ko? period1=-9999999999&period2=9999999999' url firstobs=2 dsd truncover; input date yymmdd10. +1 (4*dummy)(:$1.) ko; retain cumulative 1; cumulative=cumulative*ko/lag(ko); run; I found that cumulative without /lag(ko) recursively multiplies the historical ko values but the /lag(ko) makes everything missing. What am I doing wrong here? Updated: This is not the best but what I found. data ko2(drop=dummy); infile 'https://query1.finance.yahoo.com/v7/finance/download/ko? period1=-9999999999&period2=9999999999' url firstobs=2 dsd truncover; input date yymmdd10. +1 (4*dummy)(:$1.) ko; retain cumulative; cumulative=ifn(lag(ko)>.,cumulative*ko/lag(ko),1); run; ... View more

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. ... View more