Prophet: Support hour or minute time series in Python？

[631068264]

Prophet is a good tool.

But it only support daily data, do not support hour or minute timeseries in Python.
I hope make_future_dataframe(freq=**) cat support pandas timeseries
which can make Prophet more perfect.

And I hope I can specific the datetime column name not only ds. Or just as the index.

[maskani-moh]

@631068264

Prophet was originally thought as a tool to forecast business time series which present piece-wise trends, multiple seasonalities and which are affected by holidays. That's why frequencies are usually days, months or years.

What kind of time series are you dealing with?

[randlet]

Sub daily frequencies would be useful for us too. @631068264 The way we currently handle this is roughly as follows:

1. Do our own modeling of intra day variations of the quantity of interest to generate an intra day weighting function
2. Aggregate our data to generate a daily average time series of the quantity of interest.
3. Use Prophet to predict daily trends based on 2)
4. Use the weighting function generated in 1) to interpolate the daily Prophet model generated in 3) on an hourly basis.

not sure if that helps or not or whether such an approach would work for you.

[631068264]

Thanks your advice and answer @maskani-moh & @randlet .

This my raw data and I will use date and close as ds and y and predict the next granularity the value for close and I have different granularity such as 3min,15min,30min,1h,2h,6h,etc.

                 date     close
0 2017-06-23 09:22:00  19737.14
1 2017-06-23 09:23:00  19750.00
2 2017-06-23 09:24:00  19750.00
3 2017-06-23 09:25:00  19755.00
4 2017-06-23 09:26:00  19750.00

[hynek2001]

@maskani-moh hi, you asked, qhy hourly/minutes... i.e have router interface traffic data... data are affected by hour, by weekends, by holidays etc... my motivation is to have some 'prediction' to triger alarm, if traffic profile is unusuall

[maskani-moh]

@mraky, unfortunately, prophet was not designed for this kind of time series but more for time series showing monthly/yearly seasonality as most of the business metrics would do.

@bletham maybe you could tell us if it is planned - or desired - to enlarge the scope of prophet to a more granular level. We could work then on a PR for this feature if it's the case! 💪

[bletham]

Finer-grained time data is the top feature priority, and one of my colleagues is working on it right now. I'd expect this to be ready pretty soon (a few weeks) unless we hit unexpected roadblocks. We'll make a new branch for it and will let you know as soon as it is ready to go.

[bletham]

See also #29

[huevosabio]

Hi,

Has this been implemented and added to the main branch?

[theonewolf]

I believe this is fully supported now, I am going to try using second granularity data: https://facebook.github.io/prophet/docs/non-daily_data.html#sub-daily-data

[iamshreeram]

@theonewolf , Did you get chance to use second granular data? Can you please share your thoughts or any repo?

@631068264 , I have similar use case as you. Did you find prophet compatible? what is your suggestion for minute level data?

[bletham]

This is now fully supported. See https://facebook.github.io/prophet/docs/non-daily_data.html#sub-daily-data for an example that uses data with 5-minute granularity, but everything is the same with second-levely granularity (or even irregular spacing).

[iamshreeram]

@bletham , I believe future = m.make_future_dataframe(periods=5, freq='H') provides 5 Hourly sample. Is there a way to calculate next minute data?

Below is my ds and y

8/25/17 13:37:00 | 230
8/25/17 13:38:00 | 329
8/25/17 13:39:00 | 279
8/25/17 13:40:00 | 314
8/25/17 13:41:00 | 322
8/25/17 13:42:00 | 298
8/25/17 13:43:00 | 307
8/25/17 13:44:00 | 293
8/25/17 13:45:00 | 324
8/25/17 13:46:00 | 306
8/25/17 13:47:00 | 292

I would want to forecast below

8/25/17 13:48:00 | 302
8/25/17 13:49:00 | 298
8/25/17 13:50:00 | 309

[bletham]

future = m.make_future_dataframe(periods=5, freq='min')

You can also look at

help(m.make_future_dataframe)

to see a description of what the inputs to any function are.

[theonewolf]

@iamshreeram we did some experiments with sub-daily (second) data, but paused those experiments awhile ago. If I see any significant results or findings I will report them here.