Question: can we use the algorithm with only one trajectory?

[tinghao-kep]

Thanks for your excellent work! I just have a simple question, for some time series forecasting problems, e.g., predict the temperature in New York City, we only have one trajectory of the realization of data, unlike your example which could have as many as you want. In this scenario, is there any special considerations before using your algorithms?

Thanks!

[arunvenk]

One could create shorter time series by snipping to (ideally non-overlapping, but since weather is an IIR system, probably would keep overlap) windows to create sequences from which to learn from. The sequence length should likely be related to the time constant of the problem if the system — I would try to make the sequences to be about the length of the system’s characteristic response (how long would the effect of x_0, the first timestep affect the output).

On Wed, Dec 25, 2019 at 9:23 AM tinghao-kep ***@***.***> wrote: Thanks for your excellent work! I just have a simple question, for some time series forecasting problems, e.g., predict the temperature in New York City, we only have one trajectory of the realization of data, unlike your example which could have as many as you want. In this scenario, is there any special considerations before using your algorithms? Thanks! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=ABJWHY3QKPIOOUYCXI6NJUTQ2N3IXA5CNFSM4J7FNSU2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4ICTWQNQ>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABJWHY4CSEU2DRHPW5TC2BTQ2N3IXANCNFSM4J7FNSUQ> .

-- Sent from Gmail Mobile

[tinghao-kep]

Thanks for your prompt reply even during the Christmas, appreciated!

It is a great suggestion that the sequence length should be related with the forecast horizons. Just to clarify, let's start with a numeric example. Imagine, I have 100 days time series and want to do 10 day weather forecast. You mean ideally I should split the 100 days into 10 trajectories if assuming no overlapping?

Also, I noticed that in your demos, we do not need to specify the "y". However, sometimes I do need to use X variables to make the predictions for y as well. In this scenario, the prediction accuracy of y is more important than X. Now I am treating this equally since you codes allow arbitrary X-dimensions. Is there a better way of doing this ? Let me if you have any suggestions. Thanks!

[ovyan]

@tinghao-kep Hey! Did you succeed in using DaD for the simple data?

[tinghao-kep]

No, it does not work and the results do not make sense.

…

On Mon, May 24, 2021 at 11:26 AM Mike Ovyan ***@***.***> wrote: @tinghao-kep <https://github.com/tinghao-kep> Hey! Did you succeed in using DaD for the simple data? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANOVEYNANHS65Y7RPN5SPZ3TPJV2HANCNFSM4J7FNSUQ> .