forecasting

make_future_dataframe doesn't support regressors currently. So code like:

m = Prophet()
m.add_regressor('var')
m.fit(df)
forecasts = m.predict(m.make_future_dataframe(periods=7))

gives an error like:

ValueError: Regressor 'var' missing from dataframe when attempting to generate forecasts

I know prophet may not know what exact values to put for var in each of the rows a

The first entry is being eaten by the Differencer in its current standard setting, which may cause user frustration, especially when combined with a pipeline (which is its "typical use"), see e.g., here: alan-turing-institute/sktime#2452

We should add an NA handling parameter setting and make the default to fill in sth for the first value, e.g., a difference from an

The parameters of the likelihood model assume invalid values during training.

Description

(A clear and concise description of what the feature is.)

• util.cumsum implementation https://github.com/awslabs/gluon-ts/blob/master/src/gluonts/mx/util.py#L326 does not scale under mx.ndarray
• cumsum is 2-5 times slower than nd.cumsum under both mx.sym and mx.ndarray, and even fails for large 4-dim input

Sample test

Code

# import ...
def test_

Great first issue.

After installing NeuralProphet as developer (see CONTRIBUTING) - run pytest -v and see warning messages

Addressing these will prevent warnings becoming errors.

Is your feature request related to a problem? Please describe.

It would be nice to directly support simulating from a fitted ARIMA model, e.g. to have a simulate method to call that would delegate to statsmodels.tsa.arima.model.ARIMA.simulate. Right now, the only way I found is to use arima_res_ member of the fitted object.

Describe the solution you'd like

Class `pmdarima.arima.ar

We have a lot of antiquated docstrings that don't render well into ReadTheDocs. A kind of grunge (but incredibly useful) task would be to refactor these docstrings into proper ReadTheDocs format. This would allow us to render them effectively...

Several typos in the notebook :

• 'Would nn autorregresive '
• 'testing purporses,'
• 'will let auto_arima to handle'

🐛 Bug Report

I ran code from How To Reproduce and got strange results: fold_info column in backtest's forecasts was not added to ts[:, segment] block but concatenated separately from other segment's columns.

Expected behavior

I expected smth like

|segment |     segment_0      |     segment_1      |      segment_2     |
------------------------------------------------------

Hello, I think there is an error in the SMAPELoss Numpy implementation.

For the SMAPE function, the NumPy version always gives 100 times the PyTorch result.
Eg:
Numpy: 76.6044944858551
Pytorch: 0.7669

https://github.com/Nixtla/neuralforecast/blob/3248751ccdb1edab2ded9c14d9666eaddd1b1469/neuralforecast/losses/numpy.py#L218-L219

Since the description on the Numpy file was intended to ma