search

facebook / prophet

Tool for producing high quality forecasts for time series data that has multiple seasonality with linear or non-linear growth.
https://facebook.github.io/prophet
MIT License
18.47k stars 4.53k forks source link

fbprophet "ignores" holidays for some time series datset #2284

Open luistelmocosta opened 2 years ago

luistelmocosta commented 2 years ago

Hello, I am using a store sales time series dataset where every year, the 1st of January is a holiday and the sales are 0.

I am passing the holidays data frame provided with the dataset which contains the New Years' day tagged as a holiday. However, every time I fit the model and then run for the next year that day always has predictions above 0.

This does not happen with every store of the time series dataset. I've been trying to understand why it happens but I'm not getting anywhere.

Any idea about this kind of behavior?

Regards

vesran commented 2 years ago

Hello @luistelmocosta

Is that specific to 2017's New Year ? If that's so, I have come across the same problem. On the plot below, the predictions on New Year are always close to zero except for 2017 : image

With further inspection, this is due to a contradiction in the contribution of a weekly seasonal component and the holiday effect: the 1st January of 2017 comes out to be a Sunday while Sundays are usually known to be a day of high demand. image

FlorestNerd19 commented 1 year ago

Similar issue here. In Brazil 'dia dos finados'(20xx-11-02) usually happens during the weekdays, and as so it predicts a positive sum, instead of a negative as usual with holidays. Same thing with our independence day (20xx-11-15) image