Rolling regressions without macros
Boehmer, Broussard, and Kallunki (2002) recommend using macros to run rolling regressions. While macros make impossible tasks possible, they aren't particularly efficient. I describe here a macro-independent way of running rolling regressions, and doing similar tasks.
The task is to run rolling 24-month regressions of monthly stock excess returns on the Fama-French-Carhart factors. That is, at the end of each month, I need to:
- Get the last 24 months of excess returns for each stock, checking that there are at least, say, 12 months of data available.
- Get the Fama-French-Carhart factors.
- Run rolling regressions
I first create a dataset containing ``rankdates'', which are the date identifiers for the rolling regression. A rankdate of 31Dec2001, for instance, uses data from 31Jan2000 to 31Dec2001, inclusive.
To do this, I first get the first and last date for each permno, and then create a complete list by filling in the in-between dates. I do this because if I were to just use the permno-date pairs available in MSF, if a return is missing for a particular month, then that month will not be a rankdate, even though it might be perfectly valid to make that date a rankdate.
data firstandlastdates;
set crsp.msf(keep=permno date);
by permno; /*MSF is always sorted by permno date*/
retain firstdate;
date=intnx('month', date, 1)-1;
if first.permno then firstdate=date;
if last.permno then do;
lastdate=date;
output;
end;
run;
data permnosrankdates(rename=(date=rankdate));
set firstandlastdates;
date=firstdate;
do while(date<=lastdate);
output;
date=intnx('month', date+1, 1)-1;
end;
run;
For each rankdate, I then get the list of the 24 dates from which that rankdate will use data.
data permnosrankdates;
set permnosrankdates;
date=rankdate;
i=1;
do while(i<=24);
output;
date=intnx('month', date, 0)-1;
i=i+1;
end;
run;
permnosrankdates is a dataset that looks, in part, like this:
Obs PERMNO rankdate date i
1 10000 19851231 31DEC1985 1
2 10000 19851231 30NOV1985 2
3 10000 19851231 31OCT1985 3
4 10000 19851231 30SEP1985 4
5 10000 19851231 31AUG1985 5
6 10000 19851231 31JUL1985 6
7 10000 19851231 30JUN1985 7
8 10000 19851231 31MAY1985 8
9 10000 19851231 30APR1985 9
10 10000 19851231 31MAR1985 10
11 10000 19851231 28FEB1985 11
12 10000 19851231 31JAN1985 12
13 10000 19851231 31DEC1984 13
14 10000 19851231 30NOV1984 14
15 10000 19851231 31OCT1984 15
16 10000 19851231 30SEP1984 16
17 10000 19851231 31AUG1984 17
18 10000 19851231 31JUL1984 18
19 10000 19851231 30JUN1984 19
20 10000 19851231 31MAY1984 20
21 10000 19851231 30APR1984 21
22 10000 19851231 31MAR1984 22
23 10000 19851231 29FEB1984 23
24 10000 19851231 31JAN1984 24
25 10000 19860131 31JAN1986 1
26 10000 19860131 31DEC1985 2
27 10000 19860131 30NOV1985 3
28 10000 19860131 31OCT1985 4
29 10000 19860131 30SEP1985 5
30 10000 19860131 31AUG1985 6
We don't need to keep i, but I kept it for clarity.
Once we have this, all we need to do is merge it with the factors and the returns:
data ff;
set ff.factors_monthly(keep=date rf smb hml umd mktrf);
date=intnx('month', date, 1)-1;
run;
proc sort data=permnosrankdates;
by date permno;
data permnosrankdates;
merge permnosrankdates(in=a) ff(in=b);
by date;
if a and b;
run;
data msf;
set crsp.msf(keep=permno date ret);
where ret is not missing;
date=intnx('month', date, 1)-1;
run;
proc sort data=msf;
by date permno;
run;
/*permnosrankdates is already sorted*/
data permnosrankdates;
merge permnosrankdates(in=a) msf(in=b);
by date permno;
if a and b;
run;
Notice that I merged by date, not rankdate.
And now all that remains is to calculate excess returns and run the regressions:
data permnosrankdates;
set permnosrankdates;
exret=ret-rf;
run;
proc sort data=permnosrankdates;
by permno rankdate;
proc reg data=permnosrankdates outest=est edf;
by permno rankdate;
model exret=mktrf smb hml umd;
run;
Notice I run proc reg by rankdate, not date.
Running proc print on EST yields the following output:
Obs PERMNO rankdate _MODEL_ _TYPE_ _DEPVAR_ _RMSE_ Intercept mktrf
1 10000 19860228 MODEL1 PARMS exret . -0.2624 0.000
2 10000 19860331 MODEL1 PARMS exret . 1.9027 -32.219
3 10000 19860430 MODEL1 PARMS exret . 1.4644 -31.110
4 10000 19860531 MODEL1 PARMS exret . 4.3923 2.102
5 10000 19860630 MODEL1 PARMS exret . -14.0666 -191.074
6 10000 19860731 MODEL1 PARMS exret 0.48716 -0.1872 -4.129
7 10000 19860831 MODEL1 PARMS exret 0.34479 -0.1559 -2.784
8 10000 19860930 MODEL1 PARMS exret 0.29117 -0.1216 -0.492
9 10000 19861031 MODEL1 PARMS exret 0.25216 -0.1214 -0.499
10 10000 19861130 MODEL1 PARMS exret 0.25809 -0.0767 -1.790
11 10000 19861231 MODEL1 PARMS exret 0.27266 -0.1228 -0.709
12 10000 19870131 MODEL1 PARMS exret 0.25251 -0.1217 -0.665
13 10000 19870228 MODEL1 PARMS exret 0.23828 -0.1090 -0.784
14 10000 19870331 MODEL1 PARMS exret 0.24214 -0.1270 -1.872
15 10000 19870430 MODEL1 PARMS exret 0.22978 -0.1254 -1.942
16 10000 19870531 MODEL1 PARMS exret 0.22052 -0.1194 -1.983
17 10000 19870630 MODEL1 PARMS exret 0.22052 -0.1194 -1.983
Obs smb hml umd exret _IN_ _P_ _EDF_ _RSQ_
1 0.000 0.00 0.000 -1 0 1 0 .
2 0.000 0.00 0.000 -1 1 2 0 1.00000
3 -71.326 0.00 0.000 -1 2 3 0 1.00000
4 340.707 477.88 0.000 -1 3 4 0 1.00000
5 -884.257 -1322.51 506.315 -1 4 5 0 1.00000
6 -13.672 -16.38 7.106 -1 4 5 1 0.05020
7 -8.529 -10.71 5.470 -1 4 5 2 0.54571
8 7.327 3.24 4.754 -1 4 5 3 0.51848
9 7.303 3.22 4.751 -1 4 5 4 0.52996
10 3.312 -0.84 3.657 -1 4 5 5 0.42145
11 4.391 0.97 3.748 -1 4 5 6 0.30088
12 4.378 0.91 3.725 -1 4 5 7 0.30569
13 4.342 -0.04 3.353 -1 4 5 8 0.31433
14 0.837 -3.20 2.548 -1 4 5 9 0.26778
15 0.669 -3.35 2.531 -1 4 5 10 0.27495
16 0.604 -3.41 2.454 -1 4 5 11 0.27148
17 0.604 -3.41 2.454 -1 4 5 11 0.27148
We can tell how many observations we have for each date by looking at the _EDF_: the error degrees of freedom. We asked for this by including edf in the proc model statement. If we want to keep only regressions for which there were 12 observations or more, we keep those where _EDF_>=7.
Added by @Reeza: This is originally from:
http://pages.stern.nyu.edu/~adesouza/sasfinphd/index/node25.html
