Scalecast

Added
Changed
Fixed
- source code was using `f` instead of `self` when when calling `pop_using_criterion()`

Added
- added lstm grid in example grids
- added EarlyStopping callback functionality for the LSTM model
- added `get_expanded_lstm_grid()` to GridGenerator module which gives an example of a grid with early stopping
Changed
- changed default paramaters for the lstm model
- added `**kwargs` to the lstm model forecast function that are passed to the `fit()` function in TensorFlow, got rid of `epochs` and `batch_size` args consequently
Fixed

Added
- lstm estimator
- added `pop_using_criterion()` function
Changed
- Fixed an issue where sklearn models were being fit on the same data twice -- does not change outcomes but the models run faster now
- Output from `_scale()` function is now always a numpy matrix (could have been either that or pandas dataframe before)
- sorted `_estimators_` list
- changed the error message for when importing a grid fails to account for one other possible reason the failure occured
Fixed

Added
- Can now sort by metric value in `export_all_validation_grids_to_excel()`
Changed
Fixed
- Fixed an issue where sometimes the incorrect AR terms were being propogated for test-set evaluation only

Added
Changed
- deleted the 'scale' normalizer from the mlp grid
Fixed
- Fixed an issue with the PowerTransformer normalizer that failed because of a DivideByZero error, now defaults to a StandardScaler when this issue is encountered and logs a warning

Added
- Added `init_dates` and `levely` attributes
- Added `'Observations'` info to history and `export()`
- Added `'lvl_test_set_predictions'` to export dataframes
Changed
- Got rid of `first_obs` and `first_dates` attributes and wrote more efficient code to do what they were there for
- More information available when `__str__()` is called
Fixed
- Fixed what became an issue with the last update in which when calling `add_diffed_terms()` or `add_lagged_terms()`, the level series wasn't accurate due to how undifferencing was being executed. After examining this issue, it became evident that the previous way to undifference forecasts was less efficient than it should have been, this update fixed the issues from the last update and made the code more efficient
- Fixed an issue where AR terms were manipulating the underlying xreg structures so that each forecast were using its own test-set propogated AR values instead of the correct AR values
- Fixed the `export_Xvars_df()` method which wasn't working correctly if at least one forecast hadn't been called first