Computing cross correlation for various time stamps

[Annalena]

Hello everybody.

I have a data set which contains the following variables: date, firm and abnormal return. I attached a sample file.

I want to compute the cross correlation of the returns between the firms at each time stamp and then compute a mean cross correlation for the whole time period. I am a beginner in SAS and guess I need matrix notation but already got stuck on converting my data into a matrix.

Greatly appreciate your help!

[art297]

If you have SAS/ETS you could use: http://support.sas.com/documentation/cdl/en/etsug/63939/HTML/default/viewer.htm#etsug_timeseries_sec...

Art, CEO, AnalystFinder.com

[Annalena]

Yes, I also found this procedure and have tried it but since there is very few explanations on syntax online I don't know how to implement it.

[PeterClemmensen]

I'm not entirely sure what your desired result looks like here? 

Perhaps like this?

proc sort data= ret.abnormalreturn out=abnormalreturn;
   by date;
run;

proc transpose data=abnormalreturn out=abnormalreturn_wide(drop=date);
    by date;
    id firm;
    var ar;
run;

ods select PearsonCorr;
proc corr data=abnormalreturn_wide noprob outp=OutCorr
          nomiss;
var _numeric_;
run;

[Annalena]

Hi draycut,

i'm not sure if this does what i want since i need to calculate cross correlation and not only correlations.

I figured something out this morning and not sure if this is correct:

proc sort data=WORK.abnormalreturn out=WORK.SORTTempTableSorted;
	by Date Firm;
run;

proc transpose data=WORK.SORTTempTableSorted out=WORK.ARCor;
	var ar;
	id Firm;
	by Date;
run;

proc delete data=WORK.SORTTempTableSorted;
run;

proc timeseries data=ARcor outcrosscorr=crosscor;
	var _all_ ;
	id date interval=weekday; 
	crosscorr lag n ccf;
run;

proc means data=crosscor mean noprint;
	output out=cormean (drop=_type_ _freq_)
	mean(ccf)=roh ;
run;

Does this lead to the same output as the code you suggested? And if this is wrong and I should use your code, how would I compute a mean across all the correlations since their not in the same column?

Thanks for your support!

[StatsMan · Accepted solution]

Cross-correlations are computed over multiple time periods (ie. lags) rather than over a single time point.  Also, one typically has a single dependent variable and multiple potential inputs, and a CCF analysis is performed to help you see how correlated other potential inputs are with that dependent variable.

I wonder if a similarity analysis might be more in line with what you need…