Alright, this is a question you would best discuss with your supervisor. It looks like larger windows yield smoother backgrounds because more “smearing” is applied.
However, it seems that the numbers you change is the number of iterations, i.e. how often the algorithm runs. There is no canonical answer other than you should neither run too often (oversmooth) nor undersmooth. You seem to be testing reasonable values, and you should see some kind of optimum where changing the number of iterations doesn’t change much. This is the perfect number of iterations.
If this doesn’t happen, you will have to take this as a source of systematic uncertainties. So use a couple of reasonable values, and re-evaluate the following analysis with all of these. Use this to estimate the systematic uncertainty due to the spectrum analysis.