Multiple time series prediction

[subfeeds]

I could find some examples for time series predictions and all of them were designed to create forecasts for only one time series and I could not find any examples or workflows that generate forecasts for multiple time series (let’s say 10 time series) at once. Could someone kindly point me to some directions?

[goodvirus]

Hi @subfeeds,

it really depends what kind of forecast model you use and if all the time series should be predicted with the same model.
Fore example if you build a model which uses a decision tree and depend on lag columns, you could have all the information in a row and just have to pass multiple rows to the predictor.
Can you share an example workflow what you have so far and sample data? Maybe we can kickstart you.

Best Regards,

Paul

[subfeeds]

Hi Paul,

Thanks for the answer. For now, I’m actually just trying to run a simple model (e.g. moving average) for two time series to get the hang of which one of the loop Nodes I should use and how:

I’m not sure how it’s possible to share the workflow here.

[Thyme]

You can export workflows by right-clicking on them in the KNIME Explorer.
If you deselect the checkbox that says “Reset Workflow(s) before export” we’ll also be able to the see node output.

[subfeeds]

KNIME_project_multiple_ts.knwf (25.3 KB)

Thanks @Thyme

[goodvirus]

Hi @subfeeds,

I used your data and made a small workflow.
Essential i cut your columns and made them into rows and than did a group loop.
You also could use this to do a maschine learning model prediction.

KNIME_project_multiple_ts.knwf (67.5 KB)

Hope it helps,

Paul

[subfeeds]

Thanks Paul. That’s helpful.

I’m trying to create a workflow that takes arbitrary number of time series (e.g. 50, 100, … items) and run multiple models (e.g. Moving Average, ARIMA, MLP, etc.) for all of them and provide predictions separately. Of course it will be possible to concatenate time series the same smart way you suggested and use constant column identifier to loop over the groups; however, I believe there might be a simpler approach in the case of much more time series. I’ll appreciate it if you give a thought on that. Thanks

[goodvirus]

Hi @subfeeds,

I normally work with row data and added identifier. I build a workflow with 25.000 different time series and it worked fine.
But if your data comes with different time series always in a new column I tweaked my workflow to allow infinite different columns. There are 2 approaches.
First to cast the columns to rows with identifier (upper part of the workflow) and the second to loop per column (second part of the column).

Please be aware that we are talking about univariat time series (so only the date + target column). For multi variant time series you have to adjust the logic. Also the workflow assumes all columns have the same type.

KNIME_project_multiple_ts_infinite_columns.knwf (77.2 KB)

Hope it helps.

Best Regards,

Paul

[subfeeds]

Hi Paul, thanks a lot. That’s very helpful!

[goodvirus]

Hi @subfeeds,

your welcome. Could you mark my last post as a solution (if you have no further questions) so future forum users see the solution directly under your questions.

Thanks,

Paul