Overview
Support for Python 3.6 is dropped (2542).
Models:
- Added PyTorch implementation of the Temporal Fusion Transformer model (2536)
- Various improvements to PyTorch DeepAR (2433, 2476, 2545, 2552, 2553, 2556, 2596)
- Added wrappers for `statsforecast` models (2360, 2515, 2561)
- Added wrappers for hierarchical time series models in R (2396, 2406, 2412)
- Updated R wrappers and dockerfile (2571, 2572)
- **Important** the Naive2, Rforecast, Prophet, and Rotbaum models have been moved to `gluonts.ext` (2362, 2597)
Data:
- Improved PandasDataset: allows specifying static features as a separate dataframe,
instead of watefully replicate feature values over time. This was particularly problematic
in large datasets, such as M5. In the new setup, static features are provided via
a separate dataframe indexed by `item_id`, and the `dtype` of each column determinins
which are numerical vs categorical features, with automated detection of cardinalities
in the latter case. See the [updated tutorial notebook](https://ts.gluon.ai/stable/tutorials/data_manipulation/pandasdataframes.html) on how to use it.
Evaluation:
- New evaluation module `gluonts.ev` (2450) will gradually replace the existing
`gluonts.evaluation` as an improved, more flexible alternative.
Changelog
Breaking changes
* Remove folders of models that have moved to mx.model (2356) codingWhale13
* Remove model.common. (2358) jaheba
* Remove dataset.rolling_dataset. (2361) jaheba
* Remove DummyEstimator. (2357) jaheba
* Introduce gluonts.ext. (2362) jaheba
* Make serde.dataclass always kw-only. (2428) jaheba
* Add copy_dim to QuantileForecast, change dim method for multivariate data (2352) codingWhale13
* Include loss computation in torch DeepAR module, decouple MQF2 (2476) lostella
* Remove serde dump_code/load_code. (2482) jaheba
* Move SelfAttentionEstimator to gluonts.nursery (2534) lostella
* Require python 3.7. (2542) jaheba
* Simplify forecast.Quantile. (2544) jaheba
* Move shell related forecast classes to shell. (2547) jaheba
* Consolidate DeepNPTSEstimator (2496) lostella
* Improve PandasDataset (2573) lostella jaheba
* Simplify PandasDataset. (2583) jaheba
Major improvements / new features
* Add dict like interface for Forecast. (2384) jaheba
* Enable dropping of columns in dataset.schema.translate. (2387) jaheba
* Exposing the choice of train_sampler and validation_sampler for MQCNN and MQRNN (2381) sighellan
* Add wrapper for statsforecast models (2360) lostella
* Add dataset.schema.Schema + types. (2391) jaheba
* Add IQN implementation (1784) kashif
* Add hierarchical time series reconciliation methods from R/hts: top-down and middle-out (2396) melopeo
* Add hierarchical time series reconciliation method from R/hts: MinT (2406) melopeo
* Change schema.Type to behave like invokable types. (2443) jaheba
* Add cdf and icdf methods for StudentT distribution (2439) shchur
* Better DeepAR lags for business day frequency time series. (2433) sighellan
* Add support for feather; incl compression. (2452) jaheba
* Introduce ev module (2450) codingWhale13 jaheba
* Speed up PandasDataset further (2441) lostella
* Add Empirical Risk Minimzation (ERM) hierarchical forecasting method (2412) melopeo
* Update statsforecast model wrappers (2515) lostella
* Add nan values and explainability support for rotbaum (2537) zoolhasson
* Enable setting a custom imputation method in deepar pytorch (2545) shubhamkapoor
* Add derive_auto_fields for DeepAR PyTorch (2552) shubhamkapoor
* Add default_scale to MeanScaler and enable the option in DeepAR-PyTorch (2553) shubhamkapoor
* Add statsforecasts models (2561) melopeo
* Add TemporalFusionTransformer implementation in PyTorch (2536) shchur
* Fix r_forecast methods to work with rpy2 v3+ (2571) abdulfatir
* Updated dockerfile for R forecast models (2572) abdulfatir
* Shell: Add support for requirements.txt files. (2582) jaheba
* Expose `weight_decay` in torch TFT estimator class (2603) gorold
* Allow ReduceLROnPlateau to track val_loss when validation set is available (2614) gorold
Minor improvements / new features
* Expose SampleForecast, QuantileForecat directly in model. (2366) jaheba
* Mypy fixes (2427) jaheba
* Add nursery.pipeline. (2429) jaheba
* itertools.select. (2426) jaheba
* Add itertools.Filter. (2438) jaheba
* Add itertools.trim_nans. (2460) jaheba
* Add itertools.inverse. (2463) jaheba
* Fix: sort dataset keys in error message when importing non-existing dataset (2497) lostella
* Few shot forecasting (2517) RingoIngo
* Allow passing of additional args to dataclass. (2531) jaheba
* Simplify linear interpolation in forecast.py (2546) jaheba
* Add util.lazy_property. (2557) jaheba
* Compact PandasDataset string representation (2558) lostella
* Add default args and assertions to DeepAR pytorch module, assertions (2556) lostella
* Update MANIFEST.in. (2566) jaheba
* Add util.copy_with. (2562) jaheba
* Add missing value imputation to Seasonal Naive (2569) abdulfatir
* Implement get-item for JsonLinesFile. (2574) jaheba
* Make itertools Map/Filter dataclasses. (2579) jaheba
* Add itertools.StarMap. (2584) jaheba
* Add gluonts.maybe. (2585) jaheba
* Rework maybe. (2593) jaheba
Bug fixes
* Fix dominick dataset bug. (2364) haskarb
* Proposed fix to zero seed bug. (2379) sighellan
* Fix rotbaum random seed and num_samples argument. (2408) sighellan
* Removed unused import in test.(2409) kashif
* Hierarchical: Make sure the input S matrix is of right dtype (2415) rshyamsundar
* Speed up PandasDataset for long dataframes (2435) lostella
* Fix frequency inference in PandasDataset (2442) lostella
* Fix plotting date index bug in anomaly detection example (2446) Amrit-Bhaskar-abhask10
* Add test cases for PandasDataset, fix missing assertion (2453) lostella
* Fix MANIFEST.in (2456) lostella
* Fix pandas issue with inferring start of X frequency. (2462) jaheba
* Change default forecast_type of ND metric to median (2472) codingWhale13
* Fix: use right context in DeepVARHierarchicalEstimator (2478) c3-ziqin
* Add requirement files to MANIFEST.in (2490) jaheba
* Fix dataclass handling of member inheritance. (2492) jaheba
* Fix DateSplitter for multiples of base frequencies (2500) lostella
* Fix serde.dataclass inheritance handling. (2512) jaheba
* Fix QuantileForecast.quantile in case only mean is stored (2513) lostella
* Fix aggregate_valid for non-numerical columns (2526) lostella
* Fix dataclass eventual handling. (2530) jaheba
* Change SeasonalNaive fallback predictor to nanmean (2549) abdulfatir
* Fix: add missing params in rotbaum (2554) zoolhasson
* Add NaN validation to Evaluator (2568) abdulfatir
* Fix: avoid automatic device detection via serialized tensors when deserializing (2576) shubhamkapoor
* serde: Fix encoding of dtypes. (2586) jaheba
* Fix bug with static features in PandasDataset (2589) lostella
* Fix maybe map_or/map_or_else return types. (2588) jaheba
* Add assertion to split function ensuring valid windows (2587) MarcelK1102
* Ensure dtype on feat_time in torch DeepAR. (2596) jaheba
* Expose aggregation method in ensemble NBEATS, fix forecast shape (2598) lostella
* Fix bug with static cardinalities in `PandasDataset` (2599) lostella
* Add `gluonts.util.safe_extract` (2606) lostella
* Fix incorrect import in `tsbench`, apply latest black (2613) lostella
Documentation
* Udpate DeepAR import in README. (2359) codingWhale13
* Remove strange quoting marks from docstrings (2368) lostella
* Change 'confidence interval' to 'prediction interval' (2373) codingWhale13
* Fix use of dump_code in tutorial. (2488) jaheba
* Fix docstrings according to docformatter (2501) lostella
* Docs: Fix install instructions. (2508) jaheba
* Add examples to docstring for periods_between (2504) lostella
* Add info on how versioning works. (2529) jaheba
* Improve README example (2538) lostella
* Update REFERENCES.md dcmaddix
Test / setup changes
* Update workflow actions to latest versions (2447) lostella
* Tests: Change Python versions. (2448) jaheba
* Use ruff instead of flake8. (2485) jaheba
* Apply ruff/pyupgrade to test. (2489) jaheba
* Add smoke tests for torch models (2495) lostella
* Pin docformatter version. (2507) jaheba
* Cap numpy compatibility in mxnet extra requirements (2506) lostella
* Clean up test code for evaluator (2505) lostella
* Remove mypy plugin for dataclass. (2514) jaheba
* GH Actions: Use authenticated requests for just. (2522) jaheba
* Simplify setup.py (2525) jaheba
* Test: Only check relevant require-packages.txt for test run. (2595) jahaba
* Fix version in requirements to comply with stricter setuptools. (2604) lostella
Other
* Move NPTS back to `gluonts.model` (2597) lostella