[BUG] Inverse transform does not work on series of length < 3 about darts HOT 5 CLOSED

Hi akshayi1, thanks a lot for this detailed issue description!
You hit the nail on the head regarding the reason for this bug: The frequency was not passed to the ScalerWrapper.inverse_transform function. We just addressed this problem in PR #143 . I tested it on your code snipped and it appears to solve this issue. We also released this patch to PyPi, so you should be able to install it like this:

pip install u8darts

Please let us know if this solves your problem!

from darts.

Glad to hear it worked!

1. For longer time series, we use pandas.DatetimeIndex.inferred_freq to automatically determine the frequency. This only works for DatetimeIndex objects with a length of at least 3. So when creating a new TimeSeries instance, cases with a length shorter than 3 are handled differently. Also, we decided to warn the user when such a time series is created since it represents somewhat of an edge case. I'm sure there is still a lot of potential to improve the current approach, but this is the current setup :)
2. Your approach looks good to me! Please note that even though you're only making a single prediction, that does not necessarily mean OUTPUT_LEN has to be set to 1 for the RNNModel. Sometimes it can still useful to try a higher number since the model might learn more general trends. You could also try enhancing your univariate time series by a datetime attribute series and make it multivariate (some very basic examples: https://github.com/unit8co/darts/blob/master/examples/multivariate-examples.ipynb). Other than that, I suggest trying out our other models too, you might find one with a better fit. To find the right hyperparameters for simpler models, you can try using our backtest_gridsearch function. To get a quick overview of the performance of simpler models, our explore_models function might be interesting too (although this one is a bit experimental). But please keep in mind I am not an expert in data science.

Thanks for all your feedback!

from darts.

That works! :) Thank you so much for the quick turnaround :) Would you please be able to tell me two things?

1. Why is 3 the minimum required series length? Like, where does that requirement stem from?
2. As a general guidance, can you please tell me if I'm on the right path as far as getting the one single prediction is concerned?

I'd highly appreciate your feedback, and am looking forward to using darts. Thank you again for creating this package :)

from darts.

That makes the two of us then :P I've been a career data engineer, only just getting exposure to data science/ML. Thank you for the amazing feedback and the guidance. If possible, it'd be good to document the reason in the official documentation, as I'm sure I won't be the only one with this doubt.

Regarding your note about OUTPUT_LEN, what is the true purpose of that, if I can swap it for higher numbers, even if I want only a single prediction? Also, I don't want to hijack a bug thread to have a conversation about this, so if there's a different place we can discuss, please let me know and I'd be happy to move it there :)

Thanks again!

from darts.

Yeah, good point about the documentation, thanks! And yes, I think I should close this issue now. Feel free to also open an issue with the tag 'question' as well, or reach out to us at this address: [email protected]. Again, thanks for all the feedback!

from darts.