Scalecast

Added
- Added `MVForecaster.add_signals()` for multivariate stacking.
- Added `Forecaster.add_series()` to make multivariate RNNs possible.
- Added `return_train_only` in the `util.find_optimal_transformation()` function to add additional leaking-prevention measures to this function.
- Added padding options in `Forecaster.ingest_Xvars_df()`.
- Added `exclude_models` argument to `Reverter.fit_transform()`.
Changed
- `SeriesTransformer` no longer deletes attributes after reverting, making the object more reusable.
- `Reverter` objects always copy the passed `Forecaster` objects to make them more flexible and able to be used in more multivariate contexts.
- Added `__copy__()` and `__deepcopy__()` methods to several objects.
- Removed `Forecaster.add_diffed_terms()`.
- Changed how `Forecaster.add_lagged_terms()` deals with the introduction of N/A values -- just warn instead of chopping observations.
- Changed how `train_only` arguments are read in `SeriesTransformer` when there is no test_length in the `Forecaster` object. If test_length is 0, then `train_only` behaves as if it were False.
Fixed
- Fixed how exogenous variables are read in `util.find_optimal_lag_order()`.
- `train_only` was being ignored in `SeriesTransformer.DeseasonTransform()`.

Added
- Added functions to create expanding dynamic intervals using the conformal framework with backtesting: `util.backtest_for_resid_matrix()`, `util.get_backtest_resid_matrix()`, and `util.overwrite_forecast_intervals()`.
- Added a check for the correct estimator type with `Forecaster.auto_Xvar_select()`.
Changed
Fixed
- Fixed `Forecaster.export_Xvars_df()`, which wasn't working correctly when exogenous regressors were added with `Forecaster.ingest_Xvars_df()`.

Added
- Added tbats model.
- `MVForecaster` can now have Xvars added to it.
- Added `multiseries.line_up_dates()` function.
- New argument `test_again` added to `Forecaster.manual_forecast()` and `MVForecaster.manual_forecast()`. `auto_forecast()` in the same objects has the same new argument.
- `model` argument added to `Forecaster.save_feature_importance()` and doesn't have to be called right after a model is evaluated.
- Added `verbose` args to `cross_validate()` and `util.find_optimal_transformation()` functions to offer more transparency around these processes.
- Added `min_grid_size` argument to `Forecaster.limit_grid_size()` and `tune_test_forecast()`.
Changed
- Distinction between level/non-level is no longer tracked within `Forecaster` and `MVForecaster` objects and all methods that facilitated with that have gone away (`diff()`, `integrate()`, etc). Use `SeriesTransformer` to gain the same plus more functionality.
- Cross validation and tuning faster due to using more efficient objects.
- Cross validation has more customization available.
- RNNs and naive models can be optimized using grid search.
- All forecasting functions rewritten to be more efficient by using more efficient objects and fewer loops.
- Added `Forecaster.chop_from_front()` and `Forecaster.chop_from_back()` methods.
- Less information stored in history.
- Can only refer to series with default or user-selected names within `MVForecaster`.
- The structure of `MVForecaster` has changed. `mvf.y` is a dictionary with series in it instead of being located in `mvf.series1['y']`. The old way was to keep better track of level/nonlevel series but that no longer exists.
- test_only arg is gone in `Forecaster.manual_forecast()` in favor of `Forecaster.test()`, `MVForecaster.test()`. All out-of-sample methods facilited with `chop_from_front()` to split data and ensure no data leaks.
- `MVForecaster.add_optimizer_func()` now accepts functions.
- `MVForecaster.manual_forecast()` now accepts `Xvars` as an argument.
- `normalizer` argument considered a hyperparameter where applicable and not given its own entry in history.
- Got rid of several `Forecaster` methods that are never demonstrated in examples.
- Got rid of `Forecaster.backtest()` and `MVForecaster.backtest()`. `Pipeline.backtest()` and `MVPipeline.backtest()` are better alternatives.
Fixed
- The first element in m is taken if multiple are passed to `util.find_optimal_transformation()`. The code was taking the second.
- Fixed the diffy arg in `Forecaaster.adf_test()`.
- `cvkwargs` were not being passed to the `cross_validate()` function in `notebook.tune_test_forecast()`.
- `util.find_optimal_transformation()` was only using first value of `m` for seasonal adjustments when multiple were passed.

Added
Changed
Fixed
- Fixed the `MVPipeline.backtest()` method to return consistent results when a transformation is not taken on the `Forecaster` objects.

Added
Changed
- `util.find_optimal_transformation()` now uses pipeline backtesting, making it harder to leak data into the decision and also to make cross validation possible.
Fixed

Added
- `Forecaster`/`MVForecaster` objects pickled in now re-initiate warning logging for estimated models.
Changed
Fixed