// the find
sktime/pytorch-forecasting
Time series forecasting with PyTorch
pytorch-forecasting wraps several well-known deep learning architectures (TFT, N-BEATS, N-HiTS, DeepAR, xLSTM) behind a unified PyTorch Lightning API, with a TimeSeriesDataSet abstraction that handles the messy parts of multi-series forecasting: variable transformations, missing values, static vs time-varying covariates. It targets data scientists who want to train production-grade forecasting models without reimplementing the paper details. Now under the sktime umbrella after the original author moved on.
The TimeSeriesDataSet class genuinely earns its complexity — handling group IDs, multiple history lengths, and known-future covariates in one place saves weeks of boilerplate. TFT's attention-based interpretability (variable importance, temporal patterns) is actually implemented and usable, not just advertised. The v2 API visible in the tree (tslib_data_module, _base_model_v2) suggests the sktime team is doing real architectural work rather than just maintaining the status quo. Hyperparameter tuning via Optuna is baked in with a dedicated tuner, not an afterthought.
The v1/v2 API split is a mess in practice — you have two TimeSeriesDataSet implementations, two base model classes, two sets of tutorials, and no clear signal in the README about which one to use for new projects. The TimeSeriesDataSet requires time indices to be monotonically increasing integers, which means you must preprocess real datetime data yourself before it touches the library. Documentation quality dropped noticeably after the original maintainer handed off to sktime. The library leans hard on pandas DataFrames and will choke on datasets that don't fit in memory — there's no native streaming or chunked loading path.