input format of training and test data

[jgerlach93]

Hey,

after running the example successfully, I am trying to feed in custom data.
My feature dataset consists of many time series, of which I would only like to predict a few. The data has a frequency of 5s. All time series are non-negative, but many contain zeros. The time series that I would like to predict are all strictly larger zero.

(i) How would one adapt the code to work with 5s data? I was thinking of a quick fix to simply replace the 5s dates by daily dates? But I would still have to remove the influence of daily seasonality from the model, right?

(ii) Also, when trying to predict on just a subset of the time series by inscribing just those into the y_df, I get an assertion error in line 135 of data.py:

assert batch_y.shape[1]>=1 

(iii) I added +1 to the training and test data to prevent having zero values. Would that be ok?

I don't know whether the discussion here is still ongoing, but I would appreciate any feedback!

Best,
JC

[kdgutier]

Hi @jgerlach93 ,

In the past we added that assert to protect us against time-series with smaller length than that of inputs and outputs length.
I think the dataset is trying to feed a series that is just too short.
A simple fix would be to drop that short series. Hope it works.

In the NeuralForecast library the job of the windows creation is done by the TSDataset and TSLoader classes:
https://github.com/Nixtla/neuralforecast/blob/main/nbs/data__tsdataset.ipynb
https://github.com/Nixtla/neuralforecast/blob/main/nbs/data__tsloader.ipynb