Seasonal trend loess (STL) decomposition around holidays

Question

I am seasonally adjusting a weekly transaction time-series. The seasonal adjustment works great, except for on and around closed days. At and around these points there are serious errors. What is the proper way to decompose a time series for these localized closed days?

I can think of a couple options

1. on closed days the seasonal adjustment = the observation
2. on closed days the seasonal adjustment = trend
3. remove closed days and days around closed days

For more information, I am using decompose(type = "multiplicative"). I have tried additive, but had much worse performance.

Answer 1

Holidays are not something that can be handled by STL decomposition. Instead you should use a model that enables you to include holiday dummies. One such model is Facebook's prophet, that is able to decompose the time-series to seasonal patterns, trend, include specific dummies, and allow for structural changes by detecting change-points and using them as variables for piecewise regression.