AutoStationary

Toolkit to easily transform and inverse_transform a time-series to a stationary version and back.

Sensible parameters are automatically determined and applied with respect to the provided input parameters.

The inverse_transform() is especially practical for cases in which you have a forecasting model which was trained on transformed (e.g. differenced) labels. The output of the model in this case will also be in a transformed state, so you need to then first inverse_transform() it before you can make sense of your forecast.

I originally created AutoStationary in early 2019 when I was working on several time-series forecasting projects, and I wanted to speed up prototyping for new datasets.

The code is certainly NOT ready for production. I consider it a work-in-progress. Tested only with Python 3.10.

Implementations

Transformations:

• boxcox transform.
• differencing (1 or more orders)

Stationarity tests:

• ADF: Augmented Dickey Fuller Test
• KPSS: Kwiatkowski–Phillips–Schmidt–Shin Test.

Usage

from autostationary import AutoStationary

# transform
arr = [pandas series or numpy array]
ast = AutoStationary(arr, critical_value='5%')
transformed_arr = ast.transform()

# inverse transform, useful when your model outputs 'transformed' predictions.
transformed_preds = model_trained_on_transformed_labels.predict(X)
preds = ast.inverse_transform(transformed_preds)

# additional functions
ast.is_stationary_ADF(arr)  # returns whether arr is stationary according to ADF.
ast.is_stationary_KPSS(arr) # returns whether arr is stationary according to KPSS.
ast.summary()  # returns current state of the array in the class