I'm looking at the code for mstl decomposition and also at this mstl explanation and I see that the higher frequency seasonality (e.g. hour-of-day) is modeled first, the series is de-seasonalized and then the lower frequency (e.g. hour-of-week) is modeled.
Why is that? isn't it difficult for the hour-of-day seasonality to ignore the low weekends? why can't it be the other way around - model the hour-of-week seasonality (ignoring the daily noise by setting a low
s.window), and then de-seasonalize thus making all days the same, and then model the hour-of-day seasonality?