[Paper Review] AutoGluon-TimeSeries: AutoML for Probabilistic Time Series Forecasting

Oleksandr Shchur, Caner Türkmen+5 more|arXiv (Cornell University)|Aug 10, 2023

Forecasting Techniques and Applications20 citations

TL;DR

AutoGluon-TimeSeries (AG–TS) is an open-source AutoML library for probabilistic time series forecasting that builds ensembles of diverse models, including statistical, deep learning, and tabular forecasters, to deliver accurate point and quantile forecasts with minimal code (3 lines).

ABSTRACT

We introduce AutoGluon-TimeSeries - an open-source AutoML library for probabilistic time series forecasting. Focused on ease of use and robustness, AutoGluon-TimeSeries enables users to generate accurate point and quantile forecasts with just 3 lines of Python code. Built on the design philosophy of AutoGluon, AutoGluon-TimeSeries leverages ensembles of diverse forecasting models to deliver high accuracy within a short training time. AutoGluon-TimeSeries combines both conventional statistical models, machine-learning based forecasting approaches, and ensembling techniques. In our evaluation on 29 benchmark datasets, AutoGluon-TimeSeries demonstrates strong empirical performance, outperforming a range of forecasting methods in terms of both point and quantile forecast accuracy, and often even improving upon the best-in-hindsight combination of prior methods.

Motivation & Objective

• Democratize probabilistic time series forecasting by providing an easy-to-use AutoML tool.
• Combine diverse forecasting models (statistical, DL, tabular) through ensembling to improve accuracy and robustness.
• Enable both point and quantile forecasts with support for covariates (static and time-varying).
• Evaluate performance across a broad set of benchmark datasets to demonstrate competitive accuracy and efficiency.

Proposed method

• TimeSeriesDataFrame stores collections of univariate series with static and time-varying covariates.
• TimeSeriesPredictor defines the forecasting task, including prediction_length, quantile levels, and evaluation metric; supports cross-validation and optional hyperparameter optimization.
• Fit procedure preprocesses data, trains and evaluates multiple models, optionally tunes hyperparameters, and builds an ensemble via forward selection ( Vincentization for probabilistic forecasts).
• Three model families: local classical models (e.g., ARIMA, Theta, ETS) per-series; global models (deep learning: DeepAR, PatchTST, TFT; tabular via AutoGluon–Tabular + MLForecast); and ensembling to produce final forecasts.
• Forecasting outputs include mean (point) and quantile forecasts for the horizon, stored in a long-format DataFrame with IDs, timestamps, and predicted values.

Experimental results

Research questions

• RQ1 Can AutoGluon–TimeSeries deliver competitive point and probabilistic forecasts across diverse univariate time series datasets?
• RQ2 Does an ensemble-based AutoML approach outperform individual models and other AutoML forecasting frameworks under realistic time budgets?
• RQ3 What is the impact of including diverse model families (statistical, deep learning, tabular) and ensembling on forecasting accuracy and runtime?
• RQ4 How do presets and time-budget constraints affect accuracy and practicality for end users?

Key findings

• AG–TS outperforms competing frameworks on average for both point forecasts (MASE) and probabilistic forecasts (wQL) across 29 benchmark datasets.
• An ensemble of diverse models with forward selection achieves strong accuracy, often beating the best-in-hindsight combination.
• Three model families (statistical, deep learning, tabular) contribute meaningfully to performance, with ablations showing all are beneficial; ensembling is particularly critical.
• AG–TS maintains practical runtimes, completing training and prediction within a 4-hour budget on all datasets, with an average runtime around 33 minutes.
• Compared to AutoPyTorch, AG–TS offers competitive performance with simpler setup and often faster results on larger datasets, while AutoPyTorch may require more tuning and time.
• Ablation studies show that removing ensembling or any of the model families degrades performance, underscoring the importance of the integrated AutoML approach.