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
18.01k stars 4.48k forks source link

Outliers detection not so obvious #2473

Open Vanellope0602 opened 11 months ago

Vanellope0602 commented 11 months ago

Hi,

I am using Prophet to forecast web requests. The input meets Prophet's requirements, i.e., the ds column is the timestamp and the y column is the number of web requests in a time slice. This time series is highly periodic, with a large number of requests arriving at regular intervals, and the number of requests for the rest of the very long period being 0.

Will the prophet treat a large number of requests as outliers? Cuz I found that its prediction curve won't fit the training data points very well, it's not the same as the example graph given on the official website.

Black points are data, the blue curve is the prediction curve. I tried to adjust changepoint_prior_scale and n_changepoints but its seems not working for this problem, I want it to overfit the train data as possible as it can.

Prophet image: image

The overfit image I want:

image
AlexandroLuis commented 10 months ago

you can add add_seasonality: your_model.add_seasonality(name='monthly', period=30.5, fourier_order=6) # Adjusted Fourier order it will help with the very high and low points, adjust the fourier_order to make it even higher (try 10)