How to create rolling windows with different numbers of observations per window?

[AnjaV]

Dear SAS Community,

As a first step to perform to perform a rolling regression I would like to create rolling regression windows.

Keintz (2012) suggested following approach:

DATA rwin / view=rwin;

ws = 90;

nwin = nrecs - ws +1;

do w=1 to nwin;

do p=w to w + ws -1;

set myseries point=p nobs=nrecs;

output;

end;

end;

stop;

run;

However, I want to create annual rolling windows but with different numbers of observations per window. Window 1 must contain all firm observations from 1950-1959, window 2 all firm observation of the years 1951-1960 etc…. the last window contains all observations from 2006-2015. The number of observations differ between years.

Is it possible to solve the problem by altering the code of keintz (2012)?

Thanks very much in advance

Best regards 

Anja

[RW9]

Hi,

Can you post test data (form of a datastep) with some simple data, and what the output should look like.  If you have a date then:

data have;
  thedate="01jan1950"d; output;
  thedate="01jan1956"d; output;
  thedate="01jan1962"d; output;
  thedate="01jan1970"d; output;
run;

data want;
  set have;
  window=floor((year(thedate)-1950) / 10)+1;
run;

So you can get a window number based on year very simply with a formula.  

[AnjaV]

Hello RW9,

Thanks for your reply. This is how my data set looks like (rudimenarily)

Beob. Year FirmID Earnings LaggedEarnings 1 2 3 4 5 6 7 8 9 10 11 12 13

1970	1	100	0
1971	1	200	100
1972	1	200	200
1973	1	200	200
1974	1	200	200
1970	2	100	0
1971	2	200	100
1972	2	200	200
1970	3	100	0
1971	3	200	100
1972	3	200	200
1973	3	200	200
1974	3	200	200

And that is how the ouput should look like (but with 10 years rolling regression instead of 3 years)

Beob. Window Year FirmID Earnings LaggedEarnings 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

1	1970	1	100	0
1	1970	2	100	0
1	1970	3	100	0
1	1971	1	200	100
1	1971	2	200	100
1	1971	3	200	100
1	1972	1	200	200
1	1972	2	200	200
1	1972	3	200	200
2	1971	1	200	100
2	1971	2	200	100
2	1971	3	200	100
2	1972	1	200	200
2	1972	2	200	200
2	1972	3	200	200
2	1973	1	200	200
2	1973	3	200	2001
3	1972	1	200	200
3	1972	3	200	200
3	1973	1	200	200
3	1973	3	200	2001
3	1974	1	200	200
3	1974	3	200	200

 Thanks !!

[RW9]

Hi,

Its a good idea to post test data in the form of a datastep.

data have;
  input Year FirmID Earnings LaggedEarnings;
datalines;
1970	1	100	0
1971	1	200	100
1972	1	200	200
1973	1	200	200
1974	1	200	200
1970	2	100	0
1971	2	200	100
1972	2	200	200
1970	3	100	0
1971	3	200	100
1972	3	200	200
1973	3	200	200
1974	3	200	200
;
run;

data want (drop=inner);
  set have;
  retain window firm_id inner;
  if _n_=1 then do;
    window=1;
    firm_id=1;
    inner=1;
  end;
  else do;
    firm_id=sum(firm_id,1);
    if firm_id > 3 then do;
      firm_id=1;
      inner=sum(inner,1);
      if inner > 3 then do;
        inner=1;
        window=sum(window,1);
      end;
    end;
  end;
run;

[AnjaV]

Hi,

Thank you very much. Sorry, my statement was a misleading.

I want to divide my dataset by time-windows. The window "1970" should contain all firm data of years 1961-1970. The window "1971" should contain all firm data of years 1962-1971 and so on.... So in the end all firms of the year 1970 are listed in 10 different windows (1970-1979). 

Afterwants I want to run a pooled regression. Like:

PROC Reg Data=want Outest=Regression;

By Window;

Run; 

So that I get regression coefficients for each window.

It is not very elegant because I have a lot of doublings in my want dataset but since I am new to SAS I do not know how to code pooled rolling regression otherwise.

Thank you very much.

[RW9]

Something like (maybe a bit overcomplicated, don't really have time to look at it at the moment):

data have;
  input Year FirmID Earnings LaggedEarnings;
datalines;
1970	1	100	0
1971	1	200	100
1972	1	200	200
1973	1	200	200
1974	1	200	200
1970	2	100	0
1971	2	200	100
1972	2	200	200
1970	3	100	0
1971	3	200	100
1972	3	200	200
1973	3	200	200
1974	3	200	200
1983	3	200	200
1984	3	200	200
;
run;

data want;
  length window 8;
  set have (where=(1=0));
run;

proc sql noprint;
  select  min(YEAR),max(YEAR)
  into    :LOW,:HIGH
  from    HAVE;
quit;

%macro Get_Window;
  %local window;
  %let window=1;
  %do i=&low. %to &high. %by 10;
    data temp;
      set have (where=(&i. <= year <= %eval(&i.+10)));
      window=&window.;
    run;
    proc append base=want data=temp;
    run;
    %let window=%eval(&window.+1);
  %end;
%mend;
%Get_Window;

[AnjaV]

Hello RW9,

thanks again. Your macro runs smoothly and it divides the data into 10-years windows (e.g All firm data from 1960-1969--> window 1, data from 1970-1979 --> window 2...).

But do you maybe habe an idea how to create rolling 10-years windows? Like window 1 1960-1969; window 2 1961-197o ....?

[RW9]

Try dropping the "%by 10" part.  You will have to fiddle with it from there, I am travelling now.

[AnjaV]

RW9, thank you very much. It works

You made my day!!

[mkeintz]

I'd recommend the following sequence:

1.  Make sure your dataset is sorted by ID/YEAR
2. Make a data set VIEW containing all the needed squares and cross-products - call it  SQ_CP_BY_IDYEAR
3. Run a proc expand to generate 10-year rolling SSCP values for each firm
4. Run proc summary to pool data over all firms
5. Convert the rolling sscp into a TYPE=SSCP dataset, which PROC REG can process

The code here applies that process to sashelp.stocks (10 days instead of 10 years)

For a 10-year window this process of making 10-year rolling SSCP may not be particlularly faster than making 10-year rolling oriiginal data windows, but for daily and monthly data, where each window commonly has more records, it becomes notably superior.

proc sort data=sashelp.stocks 
   out=stocks (rename=(close=y open=ylag1 high=x1 low=x2 volume=x3 adjclose=x4));
  by stock date;
run;

/* Make a data set view of squares and cross-products */
data sqcp   /view=sqcp;
  set stocks;
  array vars {7} intercept y ylag1 x1-x4;
  retain intercept 1;

  array sqcp {7,7};
  do r=1 to 7;
    sqcp{r,r}=vars{r}**2;
    if r<7 then do c=r+1 to 7;
      sqcp{r,c}=vars{r}*vars{c};
      sqcp{c,r}=sqcp{r,c};
    end;
  end;
  drop r c;
run;

/* Make rolling sscp (SUM of squares & cross-prods) data for 10-day windows */
proc expand data=sqcp  out=rsscp_data2 (where=(intercept=10)) method=none;
  by stock;
  id date;
  convert _numeric_  / transform=(movsum 10);
run;

/* Pool sscp data over all stocks */
proc summary data=rsscp_data  nway;
  class date;
  var _numeric_;
  output out=pooled_rsscp_data (drop=_TYPE_ _FREQ_) sum=;
run;

/* Reformat the data as a TYPE=SSCP data set, ready for PROC REG */
data pooled_rolling_sscp (type=SSCP drop=sqcp:  n v);
  length _TYPE_ $8 _NAME_ $32;
  set pooled_rsscp_data;  

  array vars {7} intercept y ylag1 x1-x4;
  array sqcp {7,7};

  array vnames {7} $32 _temporary_ ;
  if _n_=1 then do v=1 to dim(vars); vnames{v}=vname(vars{v}); end;

  _TYPE_='SSCP';
  do n=1 to 7;
    _NAME_=vnames{n};
    do v=1 to 7;  vars{v}=sqcp{n,v};  end;
    output;
  end;

  _type_='N';
  _NAME_=' ';
  do v=1 to 7; vars{v}=sqcp{1,1}; end;
  output;
run;

proc reg data=pooled_rolling_sscp (obs=8);
  var y ylag1 x1-x4;
  model y=ylag1 x1-x4;
run;
quit;