Scalecast

Added
Changed
- changed how shap feature scores are sorted in reduce Xvars, no adjustment needed like with PFI
Fixed

Added
- added shap feature importances in addition to pfi by allowing user to select method = 'shap' when calling `Forecaster.save_feature_importance()`
- added shap library to dependencies list
- added `SeriesTransformer` class
- added `AnomalyDetector` class
- added function to util that breaks an `MVForecaster` class into several objects of `Forecaster` class
- can now init `Forecaster` object with `require_future_dates = False`. when using False, the object will not forecast into future dates and will not make you know values into the future for regressors passed through `Forecast.ingest_Xvars_df()`.
Changed
- took the 'per' key out of the history attribute
- changed the order of some of the source code to be more efficient (very small gains)
- changed the size of the dataset that pfi feature importance is called on to make it include all values previously seen by any given model passed to it. before, it sliced off the last couple observations only -- this was more or less a mistake but I don't expect results will be affected significantly for anyone using `reduce_Xvars()` due to how features are sorted in that function.
- changed the `Forecaster.reset()` function so that it **returns** a true copy of the initiated object.
- in `Forecaster.save_feature_importance()`, added the `on_error` arg to raise errors if the user prefers. The default is still to log errors as warnings so as not to break loops.
Fixed
- `test_only` was not working with the lstm estimator, so fixed that
- fixed an issue where the function didn't ignore the argument passed to `estimator` with `reduce_Xvars(method='l1')`

Added
Changed
Fixed
- fixed an issue where `None` wasn't being accepted in grid with `'Xvars'` as the key and using `cross_validate()` (2)

Added
Changed
- changed how the validation set length is calculated in history attributed -- given na value if cross validation used to tune models
Fixed
- fixed an issue caused by None values in hyperparam grids being changed to np.nan and therefore not accepted in some functions after cross validation has been called

Added
- added `cross_validate()` methods to `Forecaster` and `MVForecaster` objects, which can now be used for the same purposes as `tune()` but with cross validation
- added `cross_validate` as a (bool) argument to the `Forecaster.tune_test_forecast()`, `Forecaster.reduce_Xvars()`, `MVForecaster.tune_test_forecast()`, and `notebook.tune_test_forecast()` functions
- added "CrossValidated" key to history dict in `Forecaster` and `MVForecaster` objects
Changed
- if np.nan is passed as a normalizer value, it will convert to None so that it can be used
Fixed

Added
Changed
Fixed
- fixed the `set_best_model()` method in `MVForecaster`, which was broken due to not being able to parse new updates from 0.10.1