[Paper Review] GluonTS: Probabilistic Time Series Models in Python

A. Alexandrov, Konstantinos Benidis+11 more|arXiv (Cornell University)|Jun 12, 2019

Time Series Analysis and Forecasting36 references76 citations

TL;DR

GluonTS is a Python library for probabilistic, deep-learning–based time series modeling that provides components, models, data handling, evaluation, and benchmarking tools for forecasting and anomaly detection.

ABSTRACT

We introduce Gluon Time Series (GluonTS, available at https://gluon-ts.mxnet.io), a library for deep-learning-based time series modeling. GluonTS simplifies the development of and experimentation with time series models for common tasks such as forecasting or anomaly detection. It provides all necessary components and tools that scientists need for quickly building new models, for efficiently running and analyzing experiments and for evaluating model accuracy.

Motivation & Objective

• Introduce GluonTS as a toolkit for rapid development and experimentation with time series models.
• Provide components (distributions, architectures, preprocessing) to enable quick model assembly.
• Support both deep learning and probabilistic models (including state-space and Gaussian processes) for forecasting and anomaly detection.
• Demonstrate scalability from small to large datasets and reproducible experiment logging.
• Benchmark pre-bundled models on public datasets to illustrate usability and performance.

Proposed method

• Offer a modular, extensible architecture where components have clear interfaces and can be combined.
• Provide a stateless predictor API with Estimator -> Train -> Predictor workflow.
• Define a flexible distribution/output abstraction for probabilistic forecasts (Gaussian, Student-t, gamma, negative binomial, binned distributions, transformed distributions).
• Support various forecast representations (sample paths from autoregressive models, quantile forecasts) and a common evaluation interface.
• Include data I/O via DatasetRepository and a synthetic data generator, with stream-based data processing pipelines and Transformation steps.
• Implement multiple model families (state-space – including DeepState; discriminative sequence-to-sequence; auto-regressive; NPTS; DeepAR; Wavenet; Transformer) and techniques like quantile regression and neural architectures.

Experimental results

Research questions

• RQ1 How can GluonTS enable rapid construction and evaluation of probabilistic time series models across diverse tasks (forecasting, anomaly detection)?
• RQ2 What is the performance of a variety of pre-bundled models on public datasets in terms of forecast accuracy and scalability?
• RQ3 How can a unified, modular toolkit facilitate combining probabilistic distributions, neural architectures, and traditional time series components?
• RQ4 Can deep learning time series models trained on large collections of series outperform traditional local models on standard benchmarks?

Key findings

• GluonTS provides scalable, modular components that support rapid model building and evaluation for time series tasks.
• The framework includes distributions, neural architectures for sequences, and feature processing pipelines, enabling combining generative and discriminative models.
• A range of models (DeepAR, Transformer, CNN-QR, NPTS, Wavenet, DeepState) can be assembled and benchmarked on 11 public datasets.
• Experiments show neural network methods can compare favorably with ARIMA/ETS baselines across datasets, with running times suitable for practical use.