Deep learning for forecasts that admit they might be wrong

What it does

GluonTS trains deep learning estimators—like the DeepAR model shown in its examples—to forecast time series by generating probability distributions rather than single numbers. Built on PyTorch and MXNet, it ingests historical data and emits prediction intervals (the README highlights 50% and 90% bands) so you can see precisely where the model’s confidence ends.

The interesting bit

While most forecasting libraries hand you a trend line and a prayer, GluonTS treats uncertainty as the main event. That probabilistic focus has given it enough academic weight to earn a JMLR paper and recurring tutorial slots at conferences like KDD and SIGMOD.

Key highlights

• Probabilistic forecasts with built-in prediction intervals, not just point estimates.
• Backends for both PyTorch and MXNet deep learning frameworks.
• Time-series-native utilities such as PandasDataset and rolling-window train/test splitters.
• Extensive academic pedigree: JMLR publication, arXiv preprint, and a curated bibliography.
• A related zero-shot project, Chronos, is advertised but lives in a separate repository.

Caveats

• MXNet support is still advertised, despite the framework’s declining community traction.
• The prominent Chronos banner points to a separate repository, so don’t assume zero-shot models ship inside this package.

Verdict

Reach for it when your use case—demand planning, resource allocation, anomaly detection—requires knowing how wrong the model might be. If you only need a single guess and don’t care about variance, it’s overkill.