daily seasonality doesn't seem to fit properly

[sarah-mbs]

Hi everyone!

I'm facing a problem with my model. I need to predict the number of calls that my company will receive for the next month. For that, I have data from 2019 to october 2021. It is separated in intervals of 30 min each, and I have a restriction of time: i need only predictions from 8h to 23h (which is also the period of the day that i have data for). I'm quite new when it comes to forecast something, but this is my code: https://github.com/sarah-mbs/prophet.git

The problem is that i have two types of seasonality: daily and weekly. Weekly works fine, but when it comes to daily, the model predicts less calls than the actual values. I'm not really sure why this is happening. On the code you can see that i've tried different methods to solve it, but none of them works. This is the actual values vs predicted:

And these are the components:

Does anyone has any idea of how i can improve my model?

Another problem is that I can't have negative calls, but the models predicts negative values. The solution I found here was to clip those values to 0, but i'm not really sure if this was the best solution. If anyone has a different approach to this I would love to look as well.

[tcuongd]

Hmm it's also interesting that the trend is increasing even though the series looks to be fairly flat. I would try the following:

• Use growth='flat'
• Set seasonality_prior_scale=20.0 (the default is 10.0) to make the data fit more to the seasonal patterns. You can make this value higher for a stronger fit.

Regarding negative predictions, the best method for now is to clip to 0. There's a more detailed discussion of different error distributions and trend functions that try to solve this issue here: #1668