Prophet: External regressors in Prophet

[kdubovikov]

First of all, thanks for opensourcing such an awesome library!

From a distant and probably incomplete look it seems to be relatively easy to add external time series regressor to Prophet's GAM in analogy to ARIMAX models (http://robjhyndman.com/hyndsight/arimax/). This would be a very handy feature to have.

Do external regressors make sense in Prophet's mathematical framework?

[bletham]

Yes it would be very straightforward to add external regressors. The only real practical challenge is that you would need to have a forecast for the external regressors before using them to forecast the time series.

[sukwkim]

Hi @bletham and @kernelmode ,
May I ask how this task is going ? If it is possible, I want to try add bayesian linear regression for exogenous variable, spike and slab method like Bayesian structured time series model. Can I participate this task with the above approach ?

[bletham]

I expect external regressors to be done within the next week.

[bletham]

https://github.com/facebookincubator/prophet/commit/8f1607cd93d4641d690d95c5f4dd2737e8e7152f adds this in Python. See the bottom of https://github.com/facebookincubator/prophet/blob/v0.2/notebooks/seasonality_and_holiday_effects.ipynb for an example. If you have a use-case for external regressors, please install the v0.2 branch and try it out.

If someone would like to translate this commit into the R version and make a PR for it let me know, or I'll get to it next week.

[sukwkim]

@bletham awesome !!! I will translate this commit into the R and make a PR !!!

[anandgavai]

nice work @bletham, does this mean that by using "add_regressors" function we can now incorporate multiple features to predict a response variable ?

[bletham]

That is correct. The model is trend + X * beta , so the piecewise trend plus a linear component. Currently X contains features for modeling seasonality and indicator vectors for identifying holidays. The add_regressors function allows you to add whatever column to X you would like, so you can include any additional features in the linear component of the model.

[anandgavai]

@sukwkim @bletham I have problems running the v0.2 branch in python, is add_regressor available in R ?

[bletham]

Not yet on my end, @sukwkim ?

[sukwkim]

sorry for delay @anandgavai @bletham . I will add it within next week !

[cdeterman]

What is the status of this for R? I see the note by @sukwkim but no updates to prophet available. Is this just on the github perhaps?

[sukwkim]

@cdeterman Apologies, due to some personal issues, I kept delaying this. I promise I will finish it by this week.!!

[bletham]

This is available in the v0.2 release now on CRAN and pypi. Documentation is here:
https://facebookincubator.github.io/prophet/docs/seasonality_and_holiday_effects.html

If you have issues with it feel free to post here or open a new issue.

[camontanezp]

Dear all,

Does Prophet accept categorical regressors? Does it encode the regressors (e.g. one-hot) or do they need to be encoded?

Thanks

[bletham]

Extra regressors are just added directly as a feature to the linear part of the model (columns of X in X*beta). You will need to one-hot encode them yourself.

[camontanezp]

@bletham I see. Thanks!

[dsvrsec]

Can Prophet work for multivariate time series,using external regressors

[porvakanti]

Hi!
Can someone please point me to a working example of multivariate time series forecasting with fbprophet?
Thanks!

[sammo]

@porvakanti Prophet doesn't support multivariate forecasting. You'll have to predict each variable on its own.