[minor] Support data without datestamps

[SimonWittner]

🔬 Background

• Processing equidistant-data without datestamps
• Closes Non-date data: Support equidistant data without datestamps #18

🔮 Key changes

• Created a helper function to add equidistant datestamps
• Function can be manually called and adjusted to the required dates and frequency
• Function is automatically called if df consists only of column "y"

📋 Review Checklist

• I have performed a self-review of my own code.
• I have commented my code, added docstrings and data types to function definitions.
• I have added pytests to check whether my feature / fix works.

[karl-richter]

Hi @SimonWittner, I know you are just getting started with this pr - but maybe have a look at this. This will make it easier to have meaningful commit messages, pr titles and so on :)

Also, the model/metrics pytest does currently not work from forks yet. You can ignore it if that pytest fails for now.

[ourownstory]

Welcome @SimonWittner !
You picked a not so trivial issue to get started with. I look forward to hear how you plan to tackle it!
I suggest you have a chat with @judussoari as he is working on sample and datestamp processing things right now.

[SimonWittner]

Thank you for your messages! @karl-richter please let me know, if I can further improve the PR.
@ourownstory I'll get back to you!

[codecov-commenter]

Codecov Report

Merging #1188 (509c618) into main (2b303f8) will decrease coverage by 0.15%.
The diff coverage is 81.15%.

❗ Current head 509c618 differs from pull request most recent head 82ba128. Consider uploading reports for the commit 82ba128 to get more accurate results

@@            Coverage Diff             @@
##             main    #1188      +/-   ##
==========================================
- Coverage   89.88%   89.74%   -0.15%     
==========================================
  Files          38       38              
  Lines        5103     5139      +36     
==========================================
+ Hits         4587     4612      +25     
- Misses        516      527      +11     

Impacted Files	Coverage Δ
neuralprophet/data/process.py	91.96% <50.00%> (-0.60%)	⬇️
neuralprophet/forecaster.py	86.67% <79.59%> (-0.73%)	⬇️
neuralprophet/data/split.py	90.47% <80.00%> (-0.35%)	⬇️
neuralprophet/df_utils.py	95.31% <100.00%> (+0.05%)	⬆️

[ourownstory]

Thank you @SimonWittner for your contribution!

While I was thinking of adding true support for data without datestamps, which would be entail large codebase changes, your approach is a rather simple alternative that may get the job done. I think this is a good addition, so I'd like to proceed to get your contribution merge-ready.
To do so, I suggest to still do the following:

• rename mentions of 'random' to 'dummy' or 'placeholder'
• use seconds as the frequency (to avoid issues with very long time-series)
• make sure to disable auto-seasonality when your code is triggered. (as the daily/weekly/yearly seasonalities will have no meaning)
• May there be a better location (less nested and less frequently called) than prep_or_copy_df for your code to be called?
• Add to the user warning that the frequency was set to seconds
• soft suggestion: have the function only return a series of datestamps, (no passing of df), and use a more efficient implementation of adding the column to the df than concat (outside the function).

Please let me know if you disagree with my suggestions or have further ideas

[SimonWittner]

Thank you @ourownstory for your feedback! I implemented all your suggestions. Please let me know if I can further improve the PR!

[leoniewgnr]

LGTM, just change back to IDs and lets check why one test fails

[leoniewgnr]

LGTM

[ourownstory]

Good changes!

I noticed that the model is passed to the datestamp creation function.
Let us separate the disabling (editing of model config) from the timestamps creation.
So, for datestamp creation function, let us not pass the model, but only the precise args that are needed.

Further, the editing of model parameters (in a separate function), should not happen in make_future_dataframe. If seasonalities are enabled there and no datestamps present, it should create an error.

[ourownstory]

Looks good overall!

Let us separate the disabling (editing of model config) from the timestamps creation.

So, for this function, let us not pass the model, but only the precise args that are needed.

[SimonWittner]

@ourownstory in the latest commit, I removed the function call for create_dummy_datestamps from check_dataframe.
However in this commit, the create_dummy_datestamps is sometimes called in a forecaster function (for example in crossvalidation_split_df).

I think that it would be prettier to leave the function call in check_dataframe (commit 467fee9 "improved model handling_v3").

Please let me know which commit you prefer!

[SimonWittner]

Handling of data without datestamp fixed with #1432