Scalecast

Added
Changed
Fixed
- Found an issue when using `floor` in Prophet

Added
- Added the eCommerce example
- In `limit_grid_size()`, users can now set random_seed parameter for consistent results
Changed
Fixed
- Scikit-learn models were not accepting `Xvars='all'` as an arguments
- Fixed an issue causing models run at different levels to error out sometimes when plotted
- Fixed a plotting error that occured sometimes when setting models parameter to `None`

Added
- Added an option to save to png in plot(), plot_test_set(), and plot_fitted() methods using plt.savefig() from matplotlib and calling with `to_png = True`
Changed
- Made errors more descriptive, stripping out AssertionError types
Fixed
- fixed typos in doc strings

Added
Changed
- In plot() method, `models=None` is now accepted and will plot only actual values
- Example grids are modified to prevent overfitting in some models
Fixed
- Fixed the add_time_trend() method to not skip a time step in the first observation

Added
- Added a descriptive error when all_feature_info_to_excel() or all_validation_grids_to_excel() fails
Changed
- Using pd.shift() instead of np.roll() to create AR terms to avoid further issues with AR terms
- Prophet, silverkite, and ARIMA have better Xvar validation mechanisms to ensure that autoregressive terms aren't fed to them, which could cause errors and doesn't add anything to the models that isn't already built into them. Now, even if a user tries to feed AR terms only, it will pass no Xvars to these models
Fixed
- AR terms were not dropping the correct first observations before being estimated with SKLEARN models, so we fixed that but it didn't seem to make a noticeable difference in any of the examples

Added
- added reset() function that deletes all regressors and resets the object to how it was initiated
- added documentation and hints in the source code
Changed
- changed readme documentation to be more concise
Fixed
- in the documentation, it was stated that the 'scale' value passed to the 'normalize' parameter when calling manual_forecast() or auto_forecast() would use a StandardScaler from sklearn, but a Normalizer was actually being applied. Now, you can pass 'scale' to get the StandardScaler, 'normalize' to get the Normalizer, or 'minmax' to get the MinMaxScaler (unchanged from previous distributions). 'minmax' is still the default for all estimators that accept this argument