Prophet: Holidays are not being fitted anymore

[gentnm]

Hi,

I had setup fbprophet for use in a JupyterLab environment on Win10 2 weeks ago and everything was going fine. I had to do the setup again this week and now the holidays are not being recognized anymore. I can reproduce this on another machine as well. It seems there was a release of some new package version in a dependency that seems to break this, but I can't figure out which.

This is the environment.yaml I use to create the conda environment:

name: fbprophet-jupyter
channels:
    - conda-forge
    - msys2
    - defaults
dependencies:
    - python=3.7
    - cython
    - jupyterlab
    - nodejs
    - cx_oracle
    - sqlalchemy
    - plotly
    - matplotlib==3.2.2
    - fbprophet

matplotlib is set to 3.2.2 since the 3.3.0 release also slowed down plotting in fbprophet a lot. But that's another issue ;)

Any pointers or help are welcome and thank you for the awesome tool 👍

Best regards
Michael

[gentnm]

Just some update: I tried this with the official miniconda docker image (https://hub.docker.com/r/continuumio/miniconda/) as well, to be sure it is not some weird settings in my Windows/Anaconda. I can reproduce it there as well. Used the same environment file to create the environment, ran a jupyter lab and executed the following code:

%matplotlib inline
import numpy as np
import pandas as pd
from fbprophet import Prophet

dates = pd.date_range('2018-01-01', '2019-12-31')
annual = np.sin(2 * np.pi * (dates.dayofyear.values / 366 - 0.28))
noise = 15 * np.random.rand(annual.size)
def is_holiday(ds):
    date = pd.to_datetime(ds)
    return 1 if date.dayofweek==1 and date.day<8 else 0
holiday_noise = np.array([45*is_holiday(ds) for ds in dates])
data = 10 + (5 * annual) + noise + holiday_noise
s = pd.Series(data, index=dates)
df = s.to_frame().reset_index()
df = df.rename(columns={"index": "ds", 0: "y"})
df["my_holiday"] = df.ds.apply(is_holiday)

holidays = pd.DataFrame({'holiday': 'my_holiday',
                     'ds': [date for date in dates if date.dayofweek==1 and date.day<8],
                        'lower':0,
                        'upper':0})

m = Prophet(holidays=holidays)
m.fit(df)

future = m.make_future_dataframe(periods=31)

forecast = m.predict(future)

fig_forecast = m.plot(forecast)

This produced the following plot/output:

On the other hand, if I passed the holiday as an additional regressor it works:

m2 = Prophet()
m2.add_regressor("my_holiday")
m2.fit(df)

future2 = m2.make_future_dataframe(periods=31)
future2["my_holiday"] = future2.ds.apply(is_holiday)

forecast2 = m2.predict(future2)

fig_forecast2 = m2.plot(forecast2)

This confuses me a bit, since as far as I understand the holidays parameter is just a convenient wrapper around the additional regressor interface for allowing easier definition and providing lower and upper bounds.

Really hope you can help.

Edit: Tried it with a python:3.7 image and pip install jupyterlab fbprophet as well. Same results.

[bakunin75]

Can confirm this issue. It is probably related to prophet's requirements.txt using >= versioning. Some package seems to break the holidays logic.

[gentnm]

So I went through the code, only thing that could break it is pandas. So I reverted that to 1.0.5 and it is working again. Don't know what in the new release is breaking it, but at least with this we can reset it.