AutoGluonTS: Automated Time Series Forecasting

• AutoGluonTS is an open-source AutoML library for probabilistic time series forecasting that automates point and quantile predictions via model ensembling.
• It integrates classical statistical models, machine learning approaches, and deep learning architectures like DeepAR and PatchTST to achieve robust forecasting.
• The system simplifies forecasting pipelines with minimal code, automating feature engineering, model selection, and hyperparameter tuning for varied datasets.

Searching arXiv for papers on AutoGluon-TimeSeries / AutoGluonTS to ground the article in published work. AutoGluonTS, formally introduced as AutoGluon-TimeSeries, is an open-source AutoML library for probabilistic time series forecasting that emphasizes ease of use, robustness, and automated construction of accurate point and quantile forecasts (Shchur et al., 2023). Within the broader lineage of Python time-series tooling, it can be situated as a forecasting-oriented AutoML layer that combines classical statistical models, machine-learning-based forecasting approaches, and ensembling techniques, while inheriting conceptual background from GluonTS, a library for deep-learning-based time series modeling built on Apache MXNet (Alexandrov et al., 2019). Subsequent benchmark and application studies have characterized AutoGluonTS primarily as a supervised forecasting framework with automated model and hyperparameter selection, rather than as a system for unsupervised representation learning, irregularly sampled multivariate time series, or contrastive self-supervised learning (Jiao et al., 2020).

1. Origins and positioning

AutoGluonTS emerged in a research context in which GluonTS had already established a modular ecosystem for probabilistic time series models, including common tasks such as forecasting or anomaly detection, with components for data pipelines, models, distributions, training, backtesting, and evaluation (Alexandrov et al., 2019). GluonTS supports both pure deep learning models and classical probabilistic models, including RNNs, CNNs, Transformers, state space models, ARIMA, ETS, and Gaussian Processes, and exposes an Estimator/Predictor workflow for experiment-centric development (Alexandrov et al., 2019). This earlier system provides an important architectural backdrop because AutoGluonTS adopts a different emphasis: rather than primarily exposing a research framework for composing probabilistic models, it presents an AutoML library focused on robust automated forecasting with minimal user effort (Shchur et al., 2023).

The 2023 AutoGluon-TimeSeries paper defines the system as an open-source AutoML library for probabilistic time series forecasting that enables accurate point and quantile forecasts with just 3 lines of Python code (Shchur et al., 2023). Its stated design philosophy follows AutoGluon more broadly: leverage ensembles of diverse forecasting models to deliver high accuracy within a short training time (Shchur et al., 2023). This positioning distinguishes it from GluonTS as a model-development library and also from task settings such as the AutoSeries challenge, where post-hoc comparisons used AutoGluon-Tabular because no dedicated open-source AutoML framework existed for time series regression in that streaming regime (Xu et al., 2021). A plausible implication is that AutoGluonTS represents a consolidation of forecasting-specific automation capabilities that were absent in earlier tabular AutoML evaluations on time-series tasks.

Later comparative work reinforces this positioning. In the TimeAutoML paper, AutoGluonTS is described as focusing mainly on time series forecasting, supporting common deep learning architectures, and automating model and hyperparameter selection for one-step and multistep prediction, while being less focused on unsupervised tasks such as anomaly detection and clustering, irregular sampling, or contrastive self-supervised learning (Jiao et al., 2020). This characterization is consistent with its forecasting-centered API and benchmark design.

2. Forecasting formalism and supported data regime

AutoGluonTS is explicitly probabilistic. The 2023 paper frames the task as modeling the predictive distribution

$$p(y_{i, T+1:T+H} \mid y_{i, 1:T}, \mathbf{x}_{i, 1:T+H}),$$

where $\mathbf{x}$ denotes covariates and the objective is to produce both point and quantile forecasts (Shchur et al., 2023). The point forecast is typically the conditional mean,

$$\hat{y}_{i, T+1:T+H} = E_{p}\left[ y_{i, T+1:T+H} \mid y_{i, 1:T}, \mathbf{x}_{i, 1:T+H} \right],$$

while a quantile forecast $\hat{y}_{i,T+h}^q$ satisfies

$$P\left( y_{i, T+h} \leq \hat{y}_{i,T + h}^q \right) = q$$

(Shchur et al., 2023).

The system supports collections of univariate time series, or panel data, and handles static and time-varying covariates, including both known-in-advance and past-only types (Shchur et al., 2023). Predictions are returned in long-format dataframes with columns for mean and each quantile (Shchur et al., 2023). In the 2024 benchmark, AutoGluon-Timeseries is evaluated on multi-step-ahead forecasting tasks, with point forecasting only, because the benchmark restricts evaluation to point forecasts due to limitations in suitable probabilistic metrics and standard practice across the compared frameworks (Sreedhara et al., 2024). This does not contradict the library’s probabilistic scope; rather, it indicates that benchmark design may underrepresent one of its native strengths.

The relationship to GluonTS is instructive here. GluonTS formalizes forecasting over collections of time series $\{z_{i,1:T_i}\}_{i=1}^N$ with covariates $\{x_{1:T_i+\tau}\}$, learning

$$p\left(z_{i,T_i+1:T_i+\tau} \,|\, z_{i,1:T_i}, x_{1:T_i+\tau};\Phi\right)$$

(Alexandrov et al., 2019). AutoGluonTS inherits this probabilistic orientation but packages it in an AutoML workflow oriented toward production-style forecasting rather than direct model construction.

A recurring limitation in the surrounding literature is that AutoGluonTS generally expects regularly sampled, fixed-length time series and offers only limited support for irregularly sampled, variable-length multivariate time series (Jiao et al., 2020). This point should be interpreted carefully: it reflects comparative framing in an unsupervised representation-learning paper rather than a full impossibility claim, but it does delimit the problem regimes for which the system is most prominently documented.

3. Model families and ensemble mechanism

A defining property of AutoGluonTS is its combination of local statistical models and global machine-learning models within a single ensemble-based AutoML system (Shchur et al., 2023). The supported local models include ARIMA, ETS, Theta, and Seasonal Naive baselines (Shchur et al., 2023). These are fit independently per series, with StatsForecast used as the backend and Joblib used for parallelization (Shchur et al., 2023). The supported global models include deep learning and tabular machine learning models: DeepAR, PatchTST, Temporal Fusion Transformer, LightGBM via AutoGluon-Tabular, and other tree or neural regressor pipelines via MLForecast (Shchur et al., 2023).

The 2024 benchmark elaborates the model inventory used in the evaluated presets. Under the “best_quality” preset, AutoGluonTS includes a broad ensemble of deep learning, tabular, and statistical models, including DeepAR, PatchTST, Temporal Fusion Transformer, Recursive Tabular, Direct Tabular, AutoARIMA, AutoETS, and various baseline/statistical models (Sreedhara et al., 2024). The same study notes that most presets do not use hyperparameter optimization by default, in order to save time and prevent overfitting (Sreedhara et al., 2024). This suggests that the framework’s empirical strength is intended to derive primarily from model diversity and ensembling rather than exhaustive search.

The ensemble itself is constructed by forward selection. After fitting base models, AutoGluonTS employs ensemble selection via forward selection to optimize the evaluation metric over out-of-fold predictions, yielding a sparse ensemble (Shchur et al., 2023). The ensemble formulation is

$$\hat{y}_{i, T+1:T+H}^{\text{ensemble}} = \sum_{m = 1}^M w_m \cdot \hat{y}_{i, T+1:T+H}^{(m)},$$

subject to $w_m \geq 0, \sum_{m=1}^M w_m = 1$ (Shchur et al., 2023). For quantile forecasts, the ensemble is a weighted average, or Vincentization, of the individual models’ quantiles (Shchur et al., 2023).

The literature repeatedly identifies ensembling as central rather than incidental. The 2023 paper states that ensembling is a core strength and design philosophy, and ablation studies verify that removing it hurts performance substantially (Shchur et al., 2023). The 2024 benchmark likewise reports that WeightedEnsemble, RecursiveTabular, and PatchTST were most often selected or “won” within ensembles, with RecursiveTabular and PatchTST especially frequent on large datasets and long series, while AutoARIMA and Theta performed well for single or long time series (Sreedhara et al., 2024). This evidences a hybrid regime in which local statistical inductive biases remain useful for certain dataset structures, while global tabular and transformer-style models dominate many others.

4. API, automation, and workflow

AutoGluonTS is designed to minimize user intervention. The canonical interface described in the 2023 paper uses TimeSeriesDataFrame and TimeSeriesPredictor; a training pipeline can be written as:

$$\hat{y}_{i, T+1:T+H} = E_{p}\left[ y_{i, T+1:T+H} \mid y_{i, 1:T}, \mathbf{x}_{i, 1:T+H} \right],$$4

(Shchur et al., 2023)

This minimal interface is one of the central properties repeatedly emphasized in the literature. The system provides presets for balancing speed and accuracy, automatic handling of known-in-advance and past-only covariates, and robust defaults intended to serve both beginners and advanced users (Shchur et al., 2023). The robustness criterion is operationalized in a specific way: if any single model in the ensemble is trained successfully, a result is produced (Shchur et al., 2023).

Automation includes feature engineering, model and hyperparameter selection, and ensembling in application-oriented accounts. In the climatological study on forecasting seasonal extreme temperatures, AutoGluonTS is described as automating feature engineering, model/hyperparameter selection, and ensembling, while using temporal features, static station features, and exogenous inputs (Rodríguez-Bocca et al., 22 Sep 2025). The same study reports use of the "high_quality" preset, a maximum training time of 4 hours per model, a 90-day forecast horizon, and internal use of sMAPE for model ranking (Rodríguez-Bocca et al., 22 Sep 2025). This concrete deployment demonstrates that the library can absorb heterogeneous exogenous covariates such as longitude, latitude, altitude, Standardized Precipitation Indices, sea surface temperature-derived components, and geopotential-height-derived components, provided the task is formulated as a supervised forecasting problem (Rodríguez-Bocca et al., 22 Sep 2025).

From a workflow perspective, AutoGluonTS contrasts with GluonTS’s research-oriented abstractions. GluonTS exposes DatasetRepository, transformation pipelines, Forecast objects, Evaluator, and backtest_metrics, and allows users to build new models by subclassing Estimator and implementing interfaces (Alexandrov et al., 2019). AutoGluonTS instead foregrounds an AutoML predictor abstraction that hides most model-specific decisions while retaining the ability to mix local and global forecasters (Shchur et al., 2023). This suggests a division of labor: GluonTS offers a lower-level experimental substrate, whereas AutoGluonTS packages a curated subset of forecasting methodology into an automated selection-and-ensemble system.

5. Empirical performance and benchmark evidence

The principal benchmark evidence comes from the 2023 AutoGluon-TimeSeries paper and the 2024 comparative benchmark. On 29 publicly available datasets from the Monash Forecasting Repository and others, including M1, M3, M4, KDD Cup, and Electricity, AutoGluonTS was evaluated against AutoARIMA, AutoETS, AutoTheta, StatEnsemble, DeepAR, TFT, AutoPyTorch, and Seasonal Naive under comparable runtime budgets, typically 4 hours per dataset (Shchur et al., 2023). Point forecasts were evaluated by MASE and probabilistic forecasts by wQL (Shchur et al., 2023).

The MASE definition reported is

$\mathbf{x}$0

and the weighted quantile loss for quantile $\mathbf{x}$1 is

$\mathbf{x}$2

with overall wQL averaged over quantiles (Shchur et al., 2023).

The reported aggregate results are summarized below.

Evaluation	AutoGluonTS result
Point forecasting (MASE)	Champion on 19 datasets; Avg. Rank 2.08; Avg. Rescaled Error 0.073
Probabilistic forecasting (wQL)	Champion on 19 datasets; Avg. Rank 1.80; Avg. Rescaled Error 0.086
Runtime behavior	Always produced results under the 4-hour limit; mean runtime 33 minutes

(Shchur et al., 2023)

The same paper further states that AutoGluonTS outperformed a range of forecasting methods in terms of both point and quantile forecast accuracy and often even improved upon the best-in-hindsight combination of prior methods (Shchur et al., 2023). Since the detailed text also says it beat the best-in-hindsight method on 19/29 datasets, the interpretation should remain tied to those benchmark conditions rather than generalized beyond them.

The 2024 benchmark offers an independent comparison between AutoGluon-Timeseries and sktime on 36+ datasets from the Monash Time Series Forecasting Repository, covering domains such as tourism, banking, energy, economics, transportation, nature, web, sales, and health, with frequencies ranging from yearly to 4-seconds (Sreedhara et al., 2024). Under identical conditions and statistical testing using paired t-tests, Wilcoxon Signed-Rank tests, and Friedman-test-based critical difference diagrams, AutoGluonTS with a 3600-second budget ranked best for both sMAPE and MASE across most datasets (Sreedhara et al., 2024). Even with a 600-second budget, performance remained robust, although large or complex datasets sometimes limited optimal training (Sreedhara et al., 2024).

The same study reports average values summarized as sMAPE $\mathbf{x}$3 for AutoGluon-600 versus $\mathbf{x}$4 for AutoGluon-3600, and MASE $\mathbf{x}$5 for AutoGluon-600 versus $\mathbf{x}$6 for AutoGluon-3600), while cautioning that values vary by dataset and detailed tables should be consulted (Sreedhara et al., 2024). Because the paper itself emphasizes dataset dependence and de-emphasizes RMSE due to differing computation across frameworks, these aggregates should be read as benchmark-specific descriptive statistics rather than universal rankings.

6. Applications, limitations, and comparative interpretations

The application literature shows AutoGluonTS being used beyond standard benchmark datasets. In forecasting seasonal extreme temperatures over a 90-day horizon for 137 weather stations in southern South America, the target was the class of the percentage of “warm days” in the next 90 days: above normal, normal, or below normal (Rodríguez-Bocca et al., 22 Sep 2025). The task is defined through

$\mathbf{x}$7

where $\mathbf{x}$8 is a binary indicator of whether a day is warm, and forecasting seeks

$\mathbf{x}$9

with $$\hat{y}_{i, T+1:T+H} = E_{p}\left[ y_{i, T+1:T+H} \mid y_{i, 1:T}, \mathbf{x}_{i, 1:T+H} \right],$$0 (Rodríguez-Bocca et al., 22 Sep 2025). Reported multiclass AUC values are $$\hat{y}_{i, T+1:T+H} = E_{p}\left[ y_{i, T+1:T+H} \mid y_{i, 1:T}, \mathbf{x}_{i, 1:T+H} \right],$$1 for above normal, $$\hat{y}_{i, T+1:T+H} = E_{p}\left[ y_{i, T+1:T+H} \mid y_{i, 1:T}, \mathbf{x}_{i, 1:T+H} \right],$$2 for normal, and $$\hat{y}_{i, T+1:T+H} = E_{p}\left[ y_{i, T+1:T+H} \mid y_{i, 1:T}, \mathbf{x}_{i, 1:T+H} \right],$$3 for below normal (Rodríguez-Bocca et al., 22 Sep 2025). The study characterizes these results as comparable or competitive relative to the Argentinian National Meteorological Service’s operational seasonal forecast, while using a single powerful server rather than supercomputers (Rodríguez-Bocca et al., 22 Sep 2025). This suggests that AutoGluonTS can be effective as a practical forecasting back end when the prediction problem is recast into a supervised temporal classification or forecasting pipeline with rich exogenous structure.

At the same time, the surrounding literature is explicit about several limitations. In the TimeAutoML comparison, AutoGluonTS is described as having no or minimal support for unsupervised tasks such as anomaly detection and clustering, limited support for irregularly sampled, variable-length data, and no use of contrastive learning or auxiliary classification for representation learning (Jiao et al., 2020). In the AutoSeries post-hoc analysis, vanilla AutoGluon-Tabular, not AutoGluonTS, did not natively handle time series group IDs, lag generation, rolling statistics, or streaming updates, and required time-series-specific feature engineering and update logic to become competitive (Xu et al., 2021). A plausible implication is that AutoGluonTS addresses part of this gap for supervised forecasting, but not the full streaming multivariate regression regime posed by AutoSeries.

Benchmark papers also note resource-related limitations. The 2024 study reports that very massive datasets such as Kaggle Wikipedia Webtraffic stalled both AutoGluon and sktime even after 24-hour runs (Sreedhara et al., 2024). It also concludes that there is no universal dominance: for monthly data, simple Naive models can outperform AutoGluonTS on sMAPE, and Exponential Smoothing or AutoETS remain competitive in finance and economics (Sreedhara et al., 2024). The correct interpretation is therefore not that AutoGluonTS eliminates dataset dependence, but that its automated ensembles provide a strong default across diverse settings.

Later methodological comparisons sharpen this point. TSGym, a 2025 system for automated multivariate time series forecasting, explicitly contrasts its component-level model construction with AutoGluonTS’s model-level selection and ensembling, describing AutoGluonTS as selecting from a fixed set of models and hyperparameters with limited granularity for recombining deep-learning components (Liang et al., 21 Sep 2025). Adaptive pattern-specialization work likewise presents clustering-based fine-tuning and concept-drift-aware specialization as potential enhancements that AutoGluonTS could incorporate for non-stationary environments (Saadallah et al., 11 Aug 2025). These works do not invalidate AutoGluonTS’s benchmark performance; rather, they mark the boundaries of its current abstraction level.

7. Relation to GluonTS and broader AutoML directions

AutoGluonTS belongs to a broader methodological continuum in time-series automation. GluonTS provides a modular probabilistic modeling substrate with support for deep-learning and classical probabilistic models, transformation pipelines, backtesting, evaluation metrics such as quantile loss, MASE, MAPE, and sMAPE, and experiment logging for reproducibility (Alexandrov et al., 2019). AutoGluonTS builds above this ecosystem and adjacent libraries to deliver automated forecasting via ensembles of local and global models (Shchur et al., 2023). The 2023 paper notes that deep models such as DeepAR, PatchTST, and TFT are implemented via PyTorch/GluonTS, while local models rely on StatsForecast and tabular pipelines leverage AutoGluon-Tabular and MLForecast (Shchur et al., 2023). This layered architecture is significant because it makes AutoGluonTS less a single forecasting algorithm than an orchestration system over multiple forecasting paradigms.

The broader AutoML literature suggests multiple future directions. TimeAutoML emphasizes full pipeline auto-configuration for unsupervised representation learning with irregular sampling, using Thompson Sampling for module-option selection and Bayesian Optimization for hyperparameters (Jiao et al., 2020). TSGym advances a component-wise design space with meta-learning-based selection over preprocessing, encoding, architecture, and optimization choices, and empirically outperforms AutoGluonTS on several multivariate forecasting benchmarks (Liang et al., 21 Sep 2025). Pattern-specialized fine-tuning proposes clustering validation subsequences, learning specialist models, and using drift detection based on distances to cluster centroids, with improvements reported on architectures including DeepAR, MQ-CNN, DeepFactor, DeepState, and TFT as implemented in GluonTS (Saadallah et al., 11 Aug 2025).

These comparisons indicate a likely conceptual trajectory from model-level AutoML toward finer-grained, regime-aware, and transfer-oriented automation. However, within the evidence provided, AutoGluonTS remains most clearly characterized as a robust, domain-agnostic AutoML system for supervised probabilistic forecasting, distinguished by hybrid model families, sparse forward-selected ensembling, covariate support, and strong benchmark performance across diverse real-world datasets (Shchur et al., 2023, Sreedhara et al., 2024).