Prophet: Can I make a PR for cross-validation of time series?

[teramonagi]

It might be helpful for all prophet user to implement "rolling forecasting origin" type cross-validation for time series as the following link says:

• https://robjhyndman.com/hyndsight/tscv/
• https://en.wikipedia.org/wiki/Forecasting#Forecasting_accuracy

Should I implement it and make a pull request for the Prophet?

[maskani-moh]

@teramonagi

FYI, the term used for cross-validating TS is BackTesting.

Feel free to implement any feature you want and make a PR.

Your changes will be reviewed, discussed and eventually accepted or not ;)

[teramonagi]

OK, I'll give it a try :)
I know the word BackTesting, but it sounds like more financial word than time series word.

[bletham]

This is great, thanks!

[Steviey]

How can I use this new feature?

[teramonagi]

It is not released yet.
If you want to use development version, run the following commands.

$ git clone -b v0.2 https://github.com/facebookincubator/prophet.git
$ cd prophet
$ pip install .

[Steviey]

Thank you. I would need the corresponding syntax in R for install and usage.

[bletham]

I'll add it to R in the next few days.

[teramonagi]

@bletham
Let me know if you have any help!

[bletham]

This can now be installed in R using the devtools package:

devtools::install_github('facebookincubator/prophet', subdir='R', ref='v0.2')

It provides functions cross_validation and simulated_historical_forecasts. You can see the docstring for how to use the functions (?cross_validation). The output of both is a dataframe with actual values (y) and forecasted values (yhat) using a bunch of historical cutoff points. You can compare y and yhat to eachother using whichever error metric you'd like.

[bletham]

This is now in v0.2 on CRAN and pypi, with documentation here:
https://facebookincubator.github.io/prophet/docs/diagnostics.html

Thanks again!