Tutorial for Forecasting Problem

[HariWu1995]

I see that in your source code, there are 2 datasets Sunspots and Weather for forecasting problem. Could you create a demo notebook for 1 of those?
I would like to know how to build a dataset and which of your models are suitable for this problem?
Thank you in advance.

[vrodriguezf]

correct me if I'm wrong @oguiza , but any model that ends with Plus, like InceptionTimePlus is suitable for custom 3d heads, and therefore, suitable for forecasting tasks.

To build the X,y arrays for the dataset you may want to have a look at the sliding Window functionality

Best!

[oguiza]

Hi @HariWu1995,
I've been working on 1-2 forecasting tutorials with tsai for some time, but I've been very busy lately. There are also multiple variations you can implement, since it's a complex, broad area.
I'm planning to release a single-step (univariate & multivariate) forecasting notebook and then a multi-step notebook. Hopefully this will provide you some insights on how to perform this type of task with tsai.
I cannot commit though to any particular timing around this. I just need to find a few hours to finalize it.
The key to performing a forecast though is to transform the data to a supervised task. And to do that you will likely need to use the SlidingWindow functionality that @vrodriguezf mentioned before.

[farconada]

Hope to see that notebooks soon

[mandar-avhad]

HI, is the notebook for multistep forecasting been released?

[HariWu1995]

@vrodriguezf @oguiza
Thank you for your support! I have overcome this problem by my own function. No offense but your function hasn't zero-padding yet, so I need to write my own.

[oguiza]

I've moved this into discussions since it's more a request than an actual issue.

[diegoquintanav]

Hi!

Apart from using a rolling window, what are other changes that need to be made? Thanks

I've seen there is a TSForecasting transformation that is basically the same as TSRegression, and I wonder what the difference is at the implementation level. In other words, if I understand forecasting as an extrapolation task whereas regression as an interpolation task, I don't see in the examples how this is being implemented.