DataTaunew | comments | leaders | submitlogin
Announcing Prophet: A tool that provides accurate, reliable forecasting (fb.com)
30 points by danning 33 days ago | 13 comments




8 points by seanjtaylor 33 days ago | link

I'm one of the creators of Prophet. I'm happy to answer any questions people have here.

reply

2 points by don 32 days ago | link

Many thanks for releasing this! the effort to do so in both R and Python, with all the supporting documentation, is very appreciated.

Perhaps a more general or 'motivational' question - what are some common use cases for accurate events/clicks predictions? If it is just a trend that someone is interested in then an analyst can 'eyeball' it pretty well from the time series plot. I guess the need for high accuracy is driven by more than just spotting a trend?

reply

1 point by seanjtaylor 30 days ago | link

We don't really use Prophet for event prediction. It's mostly for predicting aggregates that are measured on a slower timescale (daily or weekly data). Accuracy is harder to measure for time series forecast because out of sample is tough to evaluate.

reply

1 point by saml 21 days ago | link

Are you applying any theory from the Hierarchical Temporal Memory work by Jeff Hawkins? (https://github.com/numenta/nupic/wiki/Hierarchical-Temporal-...)

It seemed to be great at forecasting periodic trends too, so that's why I wonder ...

reply

1 point by tfturing 30 days ago | link

Can it be used for causal inference (similar to counterfactual inference like CausalImpact)?

reply

1 point by fhira 31 days ago | link

What is the differentiation between Prophet's R API and Rob Hyndman's forecast package? Thanks!

reply

2 points by seanjtaylor 30 days ago | link

Good question. We tried to make it conform to a more standard fit/predict model. The main difference is that predict takes a dataframe of future dates which we provide a helper function for creating. In most of Rob Hyndman's predict functions, they take a number of periods ahead to forecast instead of an explicit set of dates.

reply

1 point by izyda 31 days ago | link

Are there plans to support covariates to the additive model?

If not, is this a feature/pull request that might be accepted?

reply

1 point by seanjtaylor 30 days ago | link

We'd love help with this. It's one of the most requested features.

reply

1 point by smartpatrol 8 days ago | link

Thanks. Great that it's using Stan.

reply

1 point by niklosb 28 days ago | link

Hi @seanjtaylor. Can you add support for multiplicative models?

I have time series that shouldn't go below zero and I couldn't find a way with your tool to bound it (similar to the way you do growth). I eventually did my own BoxCox transformation prior to modeling and had to inverse the transformation before calling the plot function. It would be great to have this built in.

reply

1 point by tatsuke 31 days ago | link

>I'm one of the creators of Prophet.

This place doesn't get much action, but I also want to say thanks. Great stuff.

reply

1 point by seanjtaylor 30 days ago | link

We're really glad to be able to share it!

reply




RSS | Announcements