NARX: Data division for multiple time series

Dear community,

I'm using NARX to predict future behaviour of a system given 6 observed time-series. Each series starts from the same initial state, but due to variing different exogeneous inputs, each series develops differently. I use catsamples (see link) to train my narx-net using all 6 series and this works quite well. Furthermore, each timeseries has a similar characheristic: low development in the first third (33%), in the middle part (33%) occurs the most changes and in the last third (33%) again less developments.

And this led my to my problem: I dont know how to divide my data. Generally, divideblock seems to be most suited for timeseries, but then the strong devleopment part will be completely coverd by training part (70%) and for validation and testing only samples remains from the last part where only small developments ocurred. In my opinion this is somehow biased, nor?

Just copy and paste them one after each other does not work either, as this causes discontinuities in the series.

I have two ideas: First idea is to use 4 complete time series for training, one for testing and one for validation. Second, use random division - but this is not recomended for timeseries, nor?

It would be great if someone has faced a similar problem or ideas how to solve this.

Best regards,

Markus

[Torsten K]

Hello Markus,

I have the exact same problem! Have you found a solution in the meantime? I would be really interested as I still don't know how to do the time series division.

Best regards

Torsten