Scalecast

Added
Changed
- `Forecaster.add_ar_terms()` now accepts collections as arguments and will add individual lags to the object according to what is passed there.
- Prophet and Silverkite models now accept direct autoregressive forecasting by passing lagged variables ('AR...') to the `Xvars` arguments in the models. As long as the lag order is the same or greater than the forecast horizon, lags are accepted.
- Cleaned up silverkite model code.
Fixed

Added
- Added `util.Forecaster_with_missing_vals()` function (66).
- Added `Forecaster.round()` method.
Changed
Fixed
- Fixed banking history for failed models. Empty dictionaries no longer saved to `Forecaster.history` nor `MVForecaster.history`. There were instances where plotting was failing as a result.

Added
- Added callbacks to the prophet model (65).
- Added `fourier_order` as a parameter in the `Forecaster.add_seasonal_regressors()` and `Forecaster.add_cycle()` methods.
Changed
Fixed

Added
Changed
Fixed
- Fixed an issue where Prophet was not adding additional regressors correctly (63).

Added
- Added the `restore_series_length()` function (62).
Changed
- Changed how in-sample metrics are evaluated. If the series length currently in the object and the predictions are different lengths, the prediction-length is truncated so that an in-sample metric can still be evaluated.
Fixed
- Fixed documentation typos.

Added
- Added the `RobustScaler` transformer and added it to the default optimal transformation search.
- Added `'robust'` as a valid normalizer argument when forecasting with scikit-learn esitmators.
Changed
Fixed