I'm currently working with a daily (business days) time series which has a monthly seasonality and an overall positive trend over the last two years. I want to estimate the error component of the time series using a STL decomposition in R. However, I am dealing with the problem that a month can have between 18-23 business days and the values of my time series are usually highest at the end of the month.
My understanding is, that for the STL decomposition to properly work, the seasonal window has to be the same number of days for each month. Therefore, my question is as follows: Is there a standard method to deal with this problem?
I've talked to a colleague who had the idea that we linearly stretch or compress each month to 20 business days. I've tried that out but I don't know if that is really a valid statistical approach. I've also googled a bit and stumbled upon a paper by the German central bank where a similar problem is discussed. Here, each month is stretched to 31 days using the "Forsythe-Malcolm-Moler algorithm" which I hadn't heard of, yet.