tinkoff-ai/etna 2.0.0

etna 2.0.0

on GitHub

Breaking changes

• Transforms now works with TSDataset instead of DataFrames:
  • Methods fit, transform and inverse_transform of transforms expect TSDataset as input
  • Transforms are not stored inside TSDataset now and should be explicitly passed into the methods fit_transform, make_future, inverse_transform
  • Forecasts from the models should be inverse transformed by the user now
• New workflow for NNs from PyTorch Forecasting, see notebook for details
• Remove some classes and methods:
  • BinsegTrendTransform - replaced with ChangePointsTrendTransform
  • sample_acf_plot, sample_pacf_plot - replaced with acf_plot
  • CatBoostModelPerSegment, CatBoostModelMultiSegment - redundant classes, CatBoostPerSegmentModel, CatBoostMultiSegmentModel are still available
  • PytorchForecastingTransform - see new workflow for NNs from PyTorch Forecasting
• Remove support of Python 3.7

Highlights:

• Add forecast decomposition for all the classical ML models, see return_components parameter in methods forecast and predict. Notebook with examples will be published soon
• Part of transforms and models are now able to work on new segments and on future data without refitting
• New backtesting strategies, see parameters refit and stride in method backtest

Full changelog:

Added

• Target components logic into AutoRegressivePipeline (#1188)
• Target components logic into HierarchicalPipeline (#1199)
• predict method into HierarchicalPipeline (#1199)
• Add target components handling in get_level_dataframe (#1179)
• Forecast decomposition for SeasonalMovingAverageModel(#1180)
• Target components logic into base classes of pipelines (#1173)
• Method predict_components for forecast decomposition in _SklearnAdapter and _LinearAdapter for linear models (#1164)
• Target components logic into base classes of models (#1158)
• Target components logic to TSDataset (#1153)
• Methods save and load to HierarchicalPipeline (#1096)
• New data access methods in TSDataset : update_columns_from_pandas, add_columns_from_pandas, drop_features (#809)
• PytorchForecastingDatasetBuiler for neural networks from Pytorch Forecasting (#971)
• New base classes for per-segment and multi-segment transforms IrreversiblePersegmentWrapper, ReversiblePersegmentWrapper, IrreversibleTransform, ReversibleTransform (#835)
• New base class for one segment transforms OneSegmentTransform (#894)
• ChangePointsLevelTransform and base classes PerIntervalModel, BaseChangePointsModelAdapter for per-interval transforms (#998)
• Method set_params to change parameters of ETNA objects (#1102)
• Function plot_forecast_decomposition (#1129)
• Method forecast_components for forecast decomposition in _TBATSAdapter (#1133)
• Methods forecast_components and predict_components for forecast decomposition in _CatBoostAdapter (#1148)
• Methods forecast_components and predict_components for forecast decomposition in _HoltWintersAdapter (#1162)
• Method predict_components for forecast decomposition in _ProphetAdapter (#1172)
• Methods forecast_components and predict_components for forecast decomposition in _SARIMAXAdapter and _AutoARIMAAdapter (#1174)
• Add refit parameter into backtest (#1159)
• Add stride parameter into backtest (#1165)
• Add optional parameter ts into forecast method of pipelines (#1071)
• Add tests on transform method of transforms on subset of segments, on new segments, on future with gap (#1094)
• Add tests on inverse_transform method of transforms on subset of segments, on new segments, on future with gap (#1127)
• In-sample prediction for BATSModel and TBATSModel (#1181)
• Method predict_components for forecast decomposition in _TBATSAdapter (#1181)
• Forecast decomposition for DeadlineMovingAverageModel(#1186)
• Prediction decomposition example into custom_transform_and_model.ipynb(#1216)

Changed

• Add optional features parameter in the signature of TSDataset.to_pandas, TSDataset.to_flatten (#809)
• Signature of the constructor of TFTModel, DeepARModel (#1110)
• Interface of Transform and PerSegmentWrapper (#835)
• Signature of TSDataset methods inverse_transform and make_future now has transforms parameter. Remove transforms and regressors updating logic from TSDataset. Forecasts from the models are not internally inverse transformed. Methods fit,transform,inverse_transform of Transform now works with TSDataset (#956)
• Create AutoBase and AutoAbstract classes, some of Auto class's logic moved there (#1114)
• Impose specific order of columns on return value of TSDataset.to_flatten (#1095)
• Add more scenarios into tests for models (#1082)
• Decouple SeasonalMovingAverageModel from PerSegmentModelMixin (#1132)
• Decouple DeadlineMovingAverageModel from PerSegmentModelMixin (#1140)
• Remove version python-3.7 from pyproject.toml, update lock (#1183)
• Bump minimum pandas version up to 1.1 (#1214)

Fixed

• Fix bug in GaleShapleyFeatureSelectionTransform with wrong number of remaining features (#1110)
• ProphetModel fails with additional seasonality set (#1157)
• Fix inference tests on new segments for DeepARModel and TFTModel (#1109)
• Fix alignment during forecasting in new NNs, add validation of context size during forecasting in new NNs, add validation of batch in MLPNet (#1108)
• Fix MeanSegmentEncoderTransform to work with subset of segments and raise error on new segments (#1104)
• Fix outliers transforms on future with gap (#1147)
• Fix SegmentEncoderTransform to work with subset of segments and raise error on new segments (#1103)
• Fix SklearnTransform in per-segment mode to work on subset of segments and raise error on new segments (#1107)
• Fix OutliersTransform and its children to raise error on new segments (#1139)
• Fix DifferencingTransform to raise error on new segments during transform and inverse_transform in inplace mode (#1141)
• Teach DifferencingTransform to inverse_transform with NaNs (#1155)
• Fixed custom_transform_and_model.ipynb(#1216)

Removed

• sample_acf_plot, sample_pacf_plot, CatBoostModelPerSegment, CatBoostModelMultiSegment (#1118)
• PytorchForecastingTransform (#971)