I have the following input:
Ticker1 | Ticker2 | Edge |
ABC | EFG | 3 |
ABC | NOM | 10 |
ABC | XYZ | 50 |
EFG | NOM | 2 |
EFG | XYZ | 0 |
EFG | QQQ | 2 |
XYZ | QQQ | 6 |
XYZ | NOM | 99 |
Which I would like to convert to an n x n matrix:
ABC | EFG | NOM | XYZ | QQQ | |
ABC | 3 | 10 | 50 | ||
EFG | 3 | 2 | 0 | 2 | |
NOM | 10 | 2 | 99 | 6 | |
XYZ | 50 | 0 | 99 | ||
QQQ | 2 | 6 |
How should I do this? the above is just to give an idea. My input dataset is:
data WORK.CLASS;
infile datalines dsd truncover;
input ticker1:$10. ticker2:$10. edge:32.;
datalines4;
ABC,EFG,3
ABC,NOM,10
ABC,XYZ,50
EFG,NOM,2
EFG,XYZ,0
EFG,QQQ,2
XYZ,QQQ,6
XYZ,NOM,99
;;;;
I cut this down to 15 obs but my dataset has 500 and the ticker set are the same for ticker1 and ticker2. Thus I would want a n x n matrix.
To clarify:
1. You say "the ticker symbols are the same for ticker1 and ticker2." Does that mean that you have a long data set that has (n x n) rows and you want to convert it to an (n x n) matrix? Or do you only have the upper triangular portion of the matrix, as shown in the little 3x3 example, which has 3 observation?
2. In the second data set, do the cell counts from the Position variable? Is the Edge variable used at all?
In your example, Ticker1 and Ticker2 do not have the same symbols. Please provide sample data that fits the problem you want to solve (or adjust your description of the problem)
1. What I meant was ticker1 and ticker2 contains tickers from a set of tickers. If you are familiar with networks, this dataset is an edge list for a network and each observation specifies the weighted connection between 2 nodes (vertices).
2. Sorry there should not be a position variable. I have removed it in the question. Regarding the second question, the Edge variable is used for the values of the matrix. So its similar to correlation between 2 tickers in the correlation matrix.
I have updated the question.
Another question:
Look at the output of the following call to PROC FREQ. Is this what you want, except you want it in a symmetric matrix in SAS/IML?
proc freq data=class;
tables ticker1*ticker2 / norow nocol nocum nopercent;
weight Position;
run;
The way to solve this problem depends on properties of the input data:
1. Are the data sorted by Ticker1 Ticker2 ?
2. Are all joint levels represented? That is, do you have an observation for every possible pair?
If the answers to (1) and (2) are both yes, then I think the following DATA step produces sample data that meets the conditions. The SAS/IML statement reads the data into the upper triangular portion of a matrix and then makes the matrix symmetric. The program uses the ROW and COL function calls, which were introduced in SAS/IML 13.1 (SAS 9.4M1).
data Have;
array c[5] $ ('A','B','C','D','E');
do i = 1 to dim(c);
Var1 = c[i];
do j = i+1 to dim(c);
Var2 = c[j];
Count = floor(10*uniform(1));
output;
end;
end;
keep Var1 Var2 Count;
run;
proc iml;
/* assume Have is sorted by (Var1, Var2) and that
all joint levels are represented (even zero counts) */
use Have;
read all var {Var1 Var2 Count};
close;
u = union(var1, var2);
n = ncol(u);
M = j(n, n, 0); /* allocate matrix of zeros */
upperTriIdx = loc(row(M) < col(M)); /* upper triangular indices */
M[ upperTriIdx ] = Count; /* assign upper triangular values */
M = M + M`; /* make symmeetric */
print M[r=u c=u];
@Rick_SAS - A minor quibble. Does not
loc(row(M) > col(M));
generate indices for lower (rather than upper) triangle?
Thanks, Mark. I have changed the inequality to match the comment.
Your example contains QQQ symbols, but your output does not contain those symbols. Please clarify.
Registration is now open for SAS Innovate 2025 , our biggest and most exciting global event of the year! Join us in Orlando, FL, May 6-9.
Sign up by Dec. 31 to get the 2024 rate of just $495.
Register now!
Learn how to run multiple linear regression models with and without interactions, presented by SAS user Alex Chaplin.
Find more tutorials on the SAS Users YouTube channel.