If you look at the help for xcorr(), you'll see how the inputs are lagged with respect to each other. If the A input is delayed by 4 samples with respect to B and you use:
[xc,lags] = xcorr(A,B,...)
You'll get the maximum at the positive lag 4.
However, if you enter
[xc,lags] = xcorr(B,A,...)
you'll get the maximum at negative lag 4.
For example:
B = randn(40,1);
% A is a delayed version of B, delayed by 4 samples
A = [zeros(4,1) ; B];
[xc,lags] = xcorr(A,B,20);
stem(lags,xc) % delay is positive
% now reverse order of inputs
[xc,lags] = xcorr(B,A,20);
figure;
stem(lags,xc) %delay is negative
In both cases, the delay is correctly indicated as 4.
