// the find
awslabs/gluonts
Probabilistic time series modeling in Python
GluonTS is Amazon's probabilistic time series forecasting library, covering everything from classical baselines to deep learning models like DeepAR and TFT. It's aimed at ML practitioners who need calibrated uncertainty estimates from their forecasts, not just point predictions. The PyTorch side is actively maintained; the MXNet side is effectively legacy.
The probabilistic forecasting primitives are well-designed — you get full predictive distributions, not just confidence intervals bolted on after the fact. The Estimator/Predictor split is clean and makes backtesting straightforward without leaking future data. Dataset handling via PandasDataset is genuinely ergonomic compared to rolling your own windowing logic. The eval framework covers proper metrics like CRPS and quantile loss, which most forecasting libraries still skip.
MXNet support is dead weight at this point — it's still in the codebase but MXNet itself is archived, so half the model zoo is effectively unmaintained. The dependency footprint is large and optional-extra-heavy; getting a working install without hitting a conflict takes trial and error. Multivariate support is limited and inconsistent across models, which is a problem the moment you have correlated series. The README prominently plugs Chronos (a separate repo) rather than explaining what GluonTS itself is good for now, which muddies the project identity.