AutoGluon-TimeSeries: AutoML for Probabilistic Time Series Forecasting

• Practitioners face a broad choice of forecasting approaches and may struggle to select methods and follow implementation and evaluation best practices.Source: §1 · p. 1
• AG–TS is an open-source Python AutoML framework for probabilistic forecasting that generates point and probabilistic forecasts for collections of univariate series.Source: §1 · p. 1
• AG–TS emphasizes ease of use and robustness, helping users with limited domain expertise produce accurate predictions with minimal coding.Source: §1 · p. 1
• The system can return a valid result when at least one model trains successfully, despite failures of individual models.Source: §1 · p. 2
• Across 29 publicly available benchmark datasets, AG–TS outperforms competing approaches in both point and probabilistic forecast accuracy.Source: §1 · p. 2

• The task is to predict the next H values for each series, with H denoting the prediction length or forecast horizon.Source: §2 · p. 2
• Forecasts condition on past series values and covariates, which may be static or time-varying and may be known in the future or only in the past.Source: §2 · p. 2
• Quantile forecasts represent possible outcomes at chosen quantile levels, whereas a conditional mean can summarize a point forecast when uncertainty is not of interest.Source: §2 · p. 2
• The setting is panel data—multiple univariate series, not multivariate forecasting—and equal series lengths are assumed for notation even though AG–TS supports differing lengths.Source: §2 · p. 2

3 AutoGluon–TimeSeries

• TimeSeriesDataFrame stores series and covariates, while TimeSeriesPredictor configures the forecast task, fits models, and returns mean and quantile forecasts.The data frame uses long format, with each row containing a series ID, timestamp, and target value.Source: §3 · p. 3, Table 1 · p. 3

3.1 Design Considerations

• AG–TS favors a broad model portfolio and ensembling over hyperparameter search, which most presets exclude to reduce training time and validation overfitting.Advanced users can still specify models, hyperparameters, or search spaces.Source: §3 · p. 3, §3.1 · p. 4
• Presets trade off training speed and accuracy, while time-based validation uses the final prediction-length window or optional multiple backtesting windows.Users can manually configure models and hyperparameters; hyperparameter optimization is optional.Source: §3.1 · p. 4

3.2 Forecasting Models

• AG–TS combines local statistical models fitted per series with global machine-learning models fitted across the dataset, including conventional methods such as ARIMA and deep-learning or tabular models.Local models must be refit for unseen series; AG–TS caches their predictions and parallelizes fitting across CPU cores.Source: §3.2 · pp. 4–5
• Global models use information across series and typically infer faster than local models, but require longer training because more parameters must be estimated.Global models also naturally handle covariates.Source: §3.2 · p. 5
• After fitting models sequentially, AG–TS uses 100 forward-selection steps to form a convex combination, tuning weights on out-of-fold forecasts for the chosen metric.For probabilistic forecasts, it averages model quantiles, a method known as Vincentization; the resulting ensemble is sparse.Source: §3.2 · p. 5

• AutoETS and AutoARIMA automatically tune parameters such as trend and seasonality for each series using an in-sample information criterion.Source: §4 · p. 5
• Later work explored combining neural-network forecasting with AutoML and automating broader pipeline stages, including preprocessing and feature engineering.Meisenbacher et al. (2022) surveyed automated forecasting pipelines.Source: §4 · p. 5
• Among open-source AutoML forecasting frameworks, AG–TS supports probabilistic forecasting and focuses on generating forecasts with few lines of code.Other listed frameworks include AutoPyTorch–Forecasting, AutoTS, and PyCaret.Source: §4 · p. 5

5.1 Setup

• The evaluation uses 29 public benchmark datasets, statistical and machine-learning baselines, and a four-hour best_quality AG–TS configuration.The benchmark set spans small datasets, a few long series, and large collections of medium-sized series.Source: §5.1 · p. 6

5.2 Forecasting Accuracy

• Some statistical frameworks fail to finish on three long-series datasets within six hours, while DeepAR fails on Web Traffic Weekly because of training numerical errors.The affected statistical models are AutoARIMA, AutoTheta, and StatEnsemble.Source: §5.2 · p. 7
• AG–TS achieved the best average rank and rescaled error for both point and probabilistic forecasts, beating the best-in-hindsight competitor on 19 of 29 datasets.The point- and probabilistic-accuracy comparisons are summarized in Tables 3 and 4.Source: §5.2 · p. 7, Table 3 · p. 7, Table 4 · p. 7

5.3 Runtime Comparison

• AG–TS completes all 29 runs in a mean of 33 minutes, avoids some extremely long local-model runtimes, and spends limited time on diverse-model ensembling rather than restricted-model tuning.AutoARIMA, AutoTheta, and StatEnsemble exceed six hours on three long-series datasets; AutoPyTorch uses its full four-hour budget.Source: §5.3 · p. 8, Figure 2 · p. 8

5.4 Ablations

• The 1-hour variant nearly matches the 4-hour version; with 10 minutes, AG–TS has the best average rank and beats the best-in-hindsight model on 14 of 29 datasets.The ablation compares reduced time limits and component variants against the baselines.Source: §5.4 · p. 9, Table 5 · p. 9
• Removing ensembling has the largest detrimental effect on accuracy, and deep-learning models are the most critical model class.The ablations also indicate that all three model classes contribute to overall performance.Source: §5.4 · p. 9

Ensembling

• Generalizing advanced ensembling strategies such as stacking to probabilistic forecasting remains an open research question.The authors identify this as an area for future development.Source: §6 · p. 9

Calibration

• Conformal prediction methods are a planned addition to address practical tasks requiring guarantees on forecast uncertainty.The authors describe these methods as one way to obtain such guarantees.Source: §6 · p. 9

New problem types

• As of v0.8, AG–TS does not support cold-start forecasting or forecast explanations.Cold-start forecasting is described as a setting with little historical data.Source: §6 · p. 10
• Combining large language models with AutoML frameworks is identified as a future-work direction.The authors mention querying AG–TS with an LLM to generate qualitative forecasts for judgemental forecasting.Source: §6 · p. 10

Scalability

• Experiments cover datasets with up to ≈107 time steps across all series; larger applications would require more efficient models and new efficient AutoML techniques.The authors note that some modern applications operate at larger scales.Source: §6 · p. 10

7 Conclusions

• AutoGluon–TimeSeries combines statistical models, deep-learning forecasting approaches, and ensembling techniques to achieve strong empirical results across benchmark datasets.The library is presented as open-source and user-friendly for probabilistic time-series forecasting.Source: §7 · p. 10
• Users can generate accurate point and quantile forecasts with just three lines of Python code.The authors present this as a way to make forecasting more accessible and efficient for a wide range of users.Source: §7 · p. 10

Broader impact

• A few lines of code can lower the barrier to forecasting for non-experts, while experienced users can use the library to design accurate forecasting pipelines.The paper connects accurate forecasts to potential impacts in renewable-energy management and inventory efficiency.Source: §8 · p. 10
• Because forecasting models capture statistical dependencies, they may mislead when used to estimate effects of actions or interventions.The authors give policy decisions as an example of decision-making contexts where overlooking forecasting limitations is dangerous.Source: §8 · p. 10

Disclosures

• The checklist reports that the work describes its limitations and potential negative societal impacts.It points to Sections 6 and 8, respectively.Source: §9 · p. 10

Reproducibility

• Supplementary materials include code, data, and instructions for reproducing the main experiments, while raw results are provided in CSV format.The checklist says the figures and tables can be reproduced from the raw data and the procedure described in the paper.Source: §9 · p. 11

Experiments

• The evaluation uses held-out final prediction windows, identical protocols across methods, and a shared four-hour runtime budget.Non-deterministic methods were repeated with five seeds, and their error metrics include means and standard deviations.Source: §9 · p. 11

Compute and tuning

• The paper reports approximately 6000 total experiment-runtime hours and describes its compute infrastructure and hyperparameter settings.The hyperparameter settings are given in Appendix A.3.Source: §9 · pp. 11–12

Human subjects

• The evaluation uses public benchmark datasets and reports no crowdsourcing or human-subjects research.The checklist marks consent and participant-risk questions as not applicable.Source: §9 · p. 12