Reconstructing time series from the decomposed trend and seasonal components from fitted prophet model

[harpreetdaurka]

Hi Prophet team, I have a question related to the one asked in issue #1740. I'm trying to make sense of the decomposed plots from a prophet model to explain it to my stakeholders. In regular techniques for seasonal decomposition (like seasonal decompose from statsmodel covered here in this article: https://towardsdatascience.com/time-series-diy-seasonal-decomposition-f0b469afed44) we get the trend, seasonal part, and errors which can be combined to recreate the original time series. However, the plots coming out of prophet are much different. The seasonality plot also changes its nature with peaks and valleys if we fit a different instance of the mode e.g. univariate vs multi-variate. The scale of the trend plot actually seems to show the smoothed overall trend of the time series over time. Is there a way for us to decompose the time series to show how much part comes from trend vs seasonality vs errors? How do I even make sense of the valleys and peaks on the yearly trend when they don't even align to actual months. Thanks in advance

[breadfan]

Hello, I have similar problem, have you solved it by any chance?