Air Quality Arena: A Large-Scale Multi-Region Ground Monitoring Dataset and Benchmark for Air Quality Forecasting with Time-Series Foundation Models

Manik Gupta; Pandarasamy Arjunan; Rishi Bharadwaj

REVIEW 4 major objections 6 minor 39 references

Air Quality Arena: A Large-Scale Multi-Region Ground Monitoring Dataset and Benchmark for Air Quality Forecasting with Time-Series Foundation Models

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that pretrained time-series foundation models, used with zero fine-tuning, beat statistical and per-station supervised baselines at 24-hour air-quality forecasting, with the top model VisionTS++ ranking first across every

desk verdict A genuinely useful multi-country air-quality benchmark with a credible zero-shot TSFM comparison, but the abstract overclaims consistency and the contamination audit is too thin to fully verify the top-model ranking. read the letter →

arxiv 2607.19381 v1 pith:KI6CQV6L submitted 2026-07-01 cs.LG

Rishi Bharadwaj , Manik Gupta , Pandarasamy Arjunan This is my paper

classification cs.LG

keywords airqualityforecastingtimeseriesfoundationmodelszero-shotbenchmarkdatasetmulti-regionmonitoringPM2.5VisionTS++cross-modallearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that pretrained time-series foundation models (TSFMs), used without any fine-tuning, forecast 24-hour air-pollution levels more accurately than classical statistical models and per-station supervised machine-learning baselines, across a newly assembled dataset of more than 14,000 station–pollutant series spanning seven countries on four continents. The authors built the Air Quality Arena dataset from six official monitoring networks, with three years of hourly readings for PM2.5, PM10, NO2, SO2, CO, and O3, and standardized filtering and imputation so the comparison is reproducible. The headline result is that VisionTS++, a model that renders time series as images and uses a vision foundation model, ranks first overall and in every country, suggesting cross-modal representations help. Why this matters: if correct, agencies can deploy accurate forecasts at new or data-poor monitoring sites without retraining, and the benchmark provides a geographically diverse testbed for a fast-moving class of models.

What carries the argument

The load-bearing object is the AQA-Data benchmark itself: 14,139 station–pollutant hourly series drawn from six national monitoring networks (EPA, CPCB, AURN, CNEMC, EEA for France/Germany, SINAICA), restricted to stations with at least 70% valid readings and no gaps over two weeks, with remaining gaps imputed in deseasonalized space via MSTL. Evaluation uses a 168-hour context window predicting the next 24 hours on a rolling 24-hour step, with scores aggregated as the geometric mean over networks of pollutant-balanced MASE and CRPS normalized by Seasonal Naive. The mechanism that carries the argument is the zero-shot setting itself: foundation models are applied with frozen weights, so any

What would settle it

A concrete check: hash the timestamp, site, and pollutant triples of AQA-Data's test-window readings and search the released pretraining corpora of the evaluated TSFMs (Chronos, Moirai, TimesFM, TiRex, VisionTS++) for exact matches; any match in the evaluation window would directly refute the zero-shot interpretation, and even near-duplicate sequences could be flagged. As a behavioral check, an independent team could train a supervised baseline on the full three years of AQA-Data and show it beats the best zero-shot TSFM, contradicting the paper's asymmetry.

Watch

Extended reading notes

Core claim

On its own data, the paper establishes that the best pretrained time-series foundation model, VisionTS++, achieves the lowest normalized MASE (0.7785) and second-lowest CRPS (0.4537) on the pollutant-balanced leaderboard, ahead of all eleven TSFMs and six classical baselines; the top TSFMs as a group beat the best supervised baseline (PatchTST, 0.8300 MASE) and all statistical baselines. The comparison is deliberately asymmetric: TSFMs see no air-quality training data, while supervised baselines are trained per pollutant per network on the first year of the same stations. The finding holds across all seven networks and all six pollutants, with the gaps to statistical baselines largest on hig

Load-bearing premise

The paper's zero-shot claim rests on the assumption that the pretrained models never saw the July 2022–June 2025 test-period readings; the authors verified this only for the public portions of a few pretraining corpora and for a subset of overlapping stations, so the check is not airtight.

Editorial extensions

If this is right

If the result holds, a single pretrained model can be dropped into any new monitoring station and produce 24-hour forecasts competitive with station-specific trained models, which is useful where historical data is sparse or unreliable.
The benchmark provides a harder, more diverse test for TSFMs than existing single-region or single-pollutant datasets; future model improvements are likely to be measured on high-pollution networks where current MASE still hovers near 0.8–0.9.
Since VisionTS++ leads on every network and on CO, NO2, PM10, and PM2.5, cross-modal architectures (image-based time-series encoders) become a plausible direction for environmental forecasting research.
The finding that supervised baselines (DLinear, LightGBM) exceed MASE 1.0 on SO2 and on CPCB suggests that on these series, even target-trained models fail to beat naive seasonality — indicating intrinsic predictability limits rather than model immaturity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If zero-shot transfer is genuine, an immediate testable extension is few-shot adaptation: fine-tuning the top TSFMs on a few weeks of CPCB data could show whether the remaining gap to perfect forecasting closes faster than training supervised models from scratch — this would separate memorization from transfer.
The cross-modal advantage invites a control experiment the paper does not run: train an equivalent pure-temporal transformer on the same image-rendered data (or ablate VisionTS++'s vision backbone) to confirm that the vision backbone, not the rendering or patch size, is responsible for the gain.
The benchmark's fixed 168-hour context and 24-hour horizon may understate or overstate TSFM strengths; testing longer horizons (48–72h) and shorter contexts could reveal whether the zero-shot edge is specific to the 24-hour horizon.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The dataset is the real contribution here. AQA-Data spans six official networks, seven countries, six pollutants, and more than 14,000 station-pollutant series over three years, collected directly from source rather than aggregated from OpenAQ. That is a substantial new resource, and the open release plus extensible codebase makes it immediately usable. The benchmark design is also thoughtful: per-network geometric mean aggregation, Seasonal Naive normalization, a realistic 168-to-24 hour forecasting setup, and per-pollutant tables that let readers see where models fail, not just where they win. The authors deserve credit for acknowledging that Kairos underperforms Seasonal Naive and that CPCB and SO2 are hard for everyone.

The central empirical claim is credible in its qualified form: the best zero-shot TSFMs (VisionTS++, TiRex, TimesFM-2.5) do beat trained supervised baselines on most networks and pollutants. But the abstract overclaims by saying TSFMs “consistently outperform classical baselines,” which is contradicted by Kairos and by multiple network-pollutant cells in the appendix. That should be softened to “top TSFMs” or “generally outperform.”

The bigger soft spot is the contamination audit in Section 3.1. The paper says overlap was confined to a subset of CNEMC and London AURN stations and that “no temporal intersection” was found, but it does not describe the method, and most TSFM pretraining corpora are not fully public. That makes the zero-shot claim conditional on an unverifiable assumption. The risk is probably moderate rather than fatal, because most of these TSFMs were pretrained on generic time series, not air-quality data, but the paper should document the audit procedure, provide per-model manifests where possible, and report sensitivity analyses that exclude the overlapping networks. Without that, the top-model ranking is not fully reproducible.

Two smaller issues: there are no confidence intervals around the leaderboard numbers, and the differences between the top three models are small enough that rankings could easily shuffle under resampling. And the post-hoc exclusion of degenerate SO2 sites (mean MASE/CRPS > 50) is a selection choice that needs a sensitivity analysis; right now it is hard to tell whether SO2 results would change materially if those sites were included.

Who is this for? Anyone building air-quality forecasting models or benchmarking TSFMs on environmental data. The paper deserves a serious referee, not a desk reject. I would send it to review and ask for a revised abstract, a real contamination audit section, uncertainty quantification on the rankings, and a robustness check on the exclusion threshold. With those changes, this becomes a standard evaluation resource.

Referee Report

4 major / 6 minor

Summary. The paper introduces Air Quality Arena (AQA), a large-scale multi-country, multi-pollutant dataset of hourly ground-monitoring air quality measurements, together with a benchmarking harness AQA-Bench. AQA-Data covers 14,139 station-pollutant series across seven countries and six pollutants over July 2022–June 2025. The benchmark evaluates 11 time-series foundation models (TSFMs) in a zero-shot setting against statistical, tree-based, and supervised deep-learning baselines for 24-hour forecasting from a 168-hour context. The headline claim is that TSFMs are effective zero-shot forecasters and consistently outperform classical baselines, with VisionTS++, a cross-modal vision-backed model, achieving the best normalized MASE. The paper also provides per-network and per-pollutant results, a preprocessing pipeline based on MSTL imputation, and a public release.

Significance. If the central claim holds, this is a useful contribution: it provides the largest geographically diverse air-quality forecasting benchmark to date, evaluates modern TSFMs in a realistic zero-shot regime, and identifies a promising cross-modal architecture. The public dataset and open evaluation harness are concrete assets that the community can build on. The paper is appropriately cautious in some places, such as noting high-pollution networks remain hardest. However, the headline claim is currently stronger than the evidence in the paper, and the zero-shot interpretation rests on a contamination audit that is not reproducible from the manuscript. The dataset itself is valuable regardless of the exact ranking, but the paper's scientific contribution is tied to the empirical comparison, so the audit and the wording of the central claim need to be fixed.

major comments (4)

[Abstract; §4.1, Table 1] The abstract and Section 4.1 state that TSFMs 'consistently outperform classical baselines.' Table 1 directly contradicts this: Kairos, a TSFM, has normalized MASE 1.0121, worse than Seasonal Naive's 1.0000, and the text itself later says Kairos is 'the only model to fall below Seasonal Naive.' The data support a weaker statement such as 'most TSFMs' or 'top-performing TSFMs.' Since this is the paper's headline claim, the wording should be corrected, and ideally supplemented with a count of station-pollutant pairs where TSFMs beat the statistical baselines.
[§3.1] The contamination assessment is load-bearing for the zero-shot claim but not verifiable. The paper states 'we found no temporal intersection between the pretraining data and AQA-Data' but gives no method, no per-model list of pretraining corpora, and no query procedure. Most evaluated TSFMs, including the top-performing VisionTS++, do not have fully public pretraining corpora. Because the headline per-country ranking in Table 48 includes CNEMC and AURN, where station-level overlap is admitted, a reader cannot rule out that part of the reported advantage comes from memorization rather than zero-shot generalization. Please provide a reproducible audit: per-model corpus provenance, date ranges, the exact overlap-check method, and a sensitivity analysis excluding all stations/networks with any overlap.
[§3.2; Appendix B] The preprocessing pipeline imputes missing values with MSTL before the chronological train/test split described in Appendix B. MSTL decomposes the full three-year series, so an imputed value in a context window can depend on observations from the evaluation period. This is temporal leakage and can inflate reported accuracy on windows containing imputed points, for all model classes. Since the benchmark claims to measure zero-shot forecasting, the imputation should be restricted to training-period data only, or the authors should show that the results are unchanged when evaluation windows containing imputed values are excluded.
[Table 1; §4.1; Appendix F] All results are point estimates without uncertainty quantification. The margin between VisionTS++ (0.7785) and TiRex (0.7825) is 0.004, and TiRex to TimesFM-2.5 (0.7831) is 0.0006, which is likely within cross-window and cross-site variation. The claims that VisionTS++ 'holds the top spot on each individual country' and that cross-modal pretraining is the reason for its success need bootstrap confidence intervals, per-network rank distributions, or a significance test. Without these, the top-model conclusion could be noise.

minor comments (6)

[Appendix C] The degenerate-site exclusion rule uses a mean MASE/CRPS threshold of 50 averaged over all models. This is a data-dependent filter applied per pollutant and is not shown to be robust. Please report how the overall leaderboard changes if these sites are retained or if the threshold is varied.
[§4.1; Table 48] 'VisionTS++ holds the top spot on each individual country' is imprecise: the evaluation is at the network level, and EEA covers two countries. The supporting table is per network, not per country.
[Table 3; Table 5; Table 29] AURN CO has only 1 site and EEA-FR CO has only 5 sites. These entries are effectively single-station or tiny-sample results and should be flagged in the main text or excluded from aggregate leaderboards, not only in Appendix A.
[Abstract; §1] The abstract says 'six monitoring networks, seven countries' but Table 2 lists seven networks (with EEA split as EEA-FR and EEA-DE) and seven countries. Clarify whether EEA is one network or two.
[§3.2; Table 1] The metric normalization is only defined in the text ('normalized by Seasonal Naive'); state explicitly in the Table 1 caption that MASE and CRPS are divided by the Seasonal Naive values, and that Seasonal Naive therefore has value 1.0000 by construction.
[Abstract] Typo: 'atAirQualityArena.github.io' is missing a space before 'at'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: this is an empirical benchmark that compares zero-shot TSFM forecasts against fitted baselines; no fitted parameter or self-citation is repackaged as a prediction.

full rationale

The paper is a dataset-and-benchmark contribution rather than a derivation. Its central claim—that zero-shot TSFMs outperform classical and supervised baselines—is supported by direct evaluation on a fixed chronological test split: TSFM checkpoints are applied without weight updates, while ML baselines are trained on the training split and statistical baselines are fitted per test window. MASE and CRPS are computed against held-out observations and normalized by Seasonal Naive, which serves as a reference baseline rather than as a fitted constant later relabeled as a finding. The aggregation procedure (mean over windows, then series, then pollutants, then geometric mean over networks) is transparent and standard and does not construct the result. No load-bearing self-citations appear: references to TIME, Gift-Eval, model papers, and toolkits are external prior work, and the authors' own prior publications are not used to justify the benchmark's conclusions. The Section 3.1 contamination audit is indeed under-specified—no method is given for checking temporal intersection and most pretraining corpora are not fully public—but that is a reproducibility and validity caveat about potential leakage, not a circular step in which the output is equivalent to an input by construction. The remaining choices, such as excluding sites with degenerate MASE/CRPS, are standard benchmark hygiene and do not force the reported ranking. Therefore the appropriate finding is no significant circularity (score 0).

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The benchmark relies on several hand-chosen thresholds and domain assumptions, but no newly postulated physical entities. The most consequential choices are the data-quality filters, the post-hoc exclusion threshold, and the unchecked contamination assumption. None are fitted to produce a particular ranking, but they do influence the composition of the dataset and the difficulty of the SO2 and CNEMC evaluations.

free parameters (5)

Data completeness threshold (70%)
Sites with less than 70% valid observations are removed, per Section 3.2; this choice shapes AQA-Data and hence every benchmark number.
Maximum gap length (2 weeks)
Sites with temporal gaps longer than two weeks are removed; this affects which series survive filtering and how much MSTL imputation is applied.
Degenerate-site exclusion threshold (mean MASE/CRPS > 50) = 50
Sites with mean MASE or CRPS > 50 across all models are excluded post hoc (Appendix C); this removes 78 sites, mostly EPA and EEA-DE SO2, materially changing the reported SO2 leaderboard.
CNEMC subsample size (200 sites per pollutant) = 200
CNEMC has roughly 1,480 sites per pollutant; a spatially stratified sample of 200 is used for evaluation, so most Chinese stations are excluded from the benchmark numbers.
Context window, horizon, step (168h, 24h, 24h)
All models predict the next 24 hours from 168 hours of context with a rolling step of 24; this design choice is noted as penalizing Kairos, which prefers longer contexts.

assumptions (4)

domain assumption Official national monitoring networks provide reliable ground-truth pollutant concentrations.
The entire dataset is built from EPA, CPCB, AURN, CNEMC, EEA, and SINAICA data, treated as ground truth (Section 3.1).
domain assumption Filtering to ≥70% completeness and ≤2-week gaps, then MSTL imputation, preserves the forecasting signal.
Section 3.2 asserts imputation in deseasonalized space is more accurate than linear interpolation; no validation of imputation fidelity against held-out real measurements is provided.
domain assumption Pretraining contamination is negligible because overlap with AQA-Data is limited to CNEMC/AURN stations with no temporal intersection.
Section 3.1 makes this claim without describing how temporal intersection was checked or how corpora were accessed; the zero-shot generalization claim depends on it.
domain assumption Excluding sites with degenerate MASE/CRPS does not bias the model comparison.
Appendix C applies exclusion independently per pollutant; this mostly affects SO2, and the paper does not show how results change if those sites are retained.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Air Quality Arena: A Large-Scale Multi-Region Ground Monitoring Dataset and Benchmark for Air Quality Forecasting with Time-Series Foundation Models." pith.science (2026). https://pith.science/paper/KI6CQV6L

@misc{pith2026260719381,
  author       = {Pith},
  title        = {Pith review of: Air Quality Arena: A Large-Scale Multi-Region Ground Monitoring Dataset and Benchmark for Air Quality Forecasting with Time-Series Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KI6CQV6L}},
  note         = {Machine review of arXiv:2607.19381}
}

read the original abstract

Air pollution causes an estimated 7.9 million premature deaths annually, making accurate forecasting a critical public health priority. Machine learning is increasingly being applied to forecast air pollution levels, yet existing benchmarks remain narrow in both geographic scope and pollutant coverage, and fail to evaluate the latest generation of time series foundation models (TSFMs) on real world, large scale data. We present Air Quality Arena (AQA), a large scale multi-country and multi-pollutant dataset (AQA-Data) and benchmark (AQA-Bench) to address this gap. AQA covers 6 major pollutants over a three year period across 7 diverse countries and 4 continents, with more than 14,000 station-pollutant series, aiming to provide a comprehensive benchmark for air quality tasks. We benchmark this dataset across 11 leading time series foundation models and classical baselines to assess performance on short-term air quality forecasting. Our results demonstrate that TSFMs are effective zero-shot forecasters and consistently outperform classical baselines, with our top-performing model employing a cross-modal architecture that leverages a vision foundation model for time series forecasting. AQA is publicly released at AirQualityArena.github.io

Figures

Figures reproduced from arXiv: 2607.19381 by the authors.

**Figure 1.** CPCB PM2.5 heatmaps across monitoring sites over time, (A) before and (B) after transformation. White regions indicate missing observations. The completeness after transformation reflects removal of sites during filtering and MSTL-based gap imputation. two weeks. This filtering was performed separately for each pollutant. Remaining gaps after site filtering were imputed using Multiple Seasonal-Trend decomposition … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 2 canonical work pages

[1]

Analysis of Air Pollution Data in India between 2015 and 2019 , volume=

Sharma, Disha and Mauzerall, Denise , year=. Analysis of Air Pollution Data in India between 2015 and 2019 , volume=. Aerosol and Air Quality Research , publisher=. doi:10.4209/aaqr.210204 , number=

work page doi:10.4209/aaqr.210204 2015
[2]

and Chen, Yue and Arnold, Steve R

Silver, Ben and Reddington, Carly L. and Chen, Yue and Arnold, Steve R. , journal =. A decade of. 2025 , doi =

2025
[3]

NeurIPS 2025 Workshop on Tackling Climate Change with Machine Learning , url=

Using Time Series Foundation Models for Atmospheric CO2 Concentration Forecasting , author=. NeurIPS 2025 Workshop on Tackling Climate Change with Machine Learning , url=

2025
[4]

Continuous Ambient Air Quality Monitoring (
[5]

Carslaw and Karl Ropkins , keywords =

David C. Carslaw and Karl Ropkins , keywords =. openair — An R package for air quality data analysis , journal =. 2012 , issn =. doi:https://doi.org/10.1016/j.envsoft.2011.09.008 , url =

work page doi:10.1016/j.envsoft.2011.09.008 2012
[6]

China Air Quality Historical Data , author =
[7]

Air Quality Download Service , author =
[8]

Wilson, Robin , url =

Show all 39 references

[9]

Automatic Urban and Rural Network (
[10]

International Journal on Recent and Innovation Trends in Computing and Communication , volume =

Comparison of Imputation Methods for Univariate Time Series , author =. International Journal on Recent and Innovation Trends in Computing and Communication , volume =. 2023 , doi =

2023
[11]

Atmosphere , volume =

Mind the Large Gap: Novel Algorithm Using Seasonal Decomposition and Elastic Net Regression to Impute Large Intervals of Missing Data in Air Quality Data , author =. Atmosphere , volume =. 2023 , doi =

2023
[12]

and Cleveland, William S

Cleveland, Robert B. and Cleveland, William S. and McRae, Jean E. and Terpenning, Irma , title =. Journal of Official Statistics , volume =. 1990 , publisher =

1990
[13]

Olivares , title =

Azul Garza and Max Mergenthaler Canseco and Cristian Challú and Kin G. Olivares , title =. 2022 , howpublished=

2022
[14]

2024 , eprint=

GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation , author=. 2024 , eprint=

2024
[15]

arXiv preprint arXiv:2602.12147 , year=

It's TIME: Towards the Next Generation of Time Series Forecasting Benchmarks , author=. arXiv preprint arXiv:2602.12147 , year=

arXiv
[16]

OpenAQ , howpublished =
[17]

World Air Quality Index , howpublished =
[18]

and Stomberg, T

Betancourt, C. and Stomberg, T. and Roscher, R. and Schultz, M. G. and Stadtler, S. , TITLE =. Earth System Science Data , VOLUME =. 2021 , NUMBER =

2021
[19]

arXiv preprint arXiv:2511.11698 , year=

Moirai 2.0: When less is more for time series forecasting , author=. arXiv preprint arXiv:2511.11698 , year=
[20]

Forty-first International Conference on Machine Learning , year=

Unified Training of Universal Time Series Forecasting Transformers , author=. Forty-first International Conference on Machine Learning , year=
[21]

Transactions on Machine Learning Research , issn=

Chronos: Learning the Language of Time Series , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

2024
[22]

arXiv preprint arXiv:2510.15821 , year=

Chronos-2: From univariate to universal forecasting , author=. arXiv preprint arXiv:2510.15821 , year=

arXiv
[23]

arXiv preprint arXiv:2508.04379 , year=

VisionTS++: Cross-Modal Time Series Foundation Model with Continual Pre-trained Vision Backbones , author=. arXiv preprint arXiv:2508.04379 , year=
[24]

The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

TiRex: Zero-Shot Forecasting Across Long and Short Horizons with Enhanced In-Context Learning , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=
[25]

Forty-second International Conference on Machine Learning , year=

Sundial: A Family of Highly Capable Time Series Foundation Models , author=. Forty-second International Conference on Machine Learning , year=
[26]

2026 , eprint=

Kairos: Toward Adaptive and Parameter-Efficient Time Series Foundation Models , author=. 2026 , eprint=

2026
[27]

2024 , eprint=

A decoder-only foundation model for time-series forecasting , author=. 2024 , eprint=

2024
[28]

State of Global Air 2025: A Report on Air Pollution and Its Role in the World's Leading Causes of Death , year =

2025
[29]

Air Pollution , howpublished =
[30]

Effective PM2.5 concentration forecasting based on multiple spatial-temporal GNN for areas without monitoring stations , journal =

I-Fang Su and Yu-Chi Chung and Chiang Lee and Pin-Man Huang , keywords =. Effective PM2.5 concentration forecasting based on multiple spatial-temporal GNN for areas without monitoring stations , journal =. 2023 , issn =. doi:https://doi.org/10.1016/j.eswa.2023.121074 , url =

2023
[31]

AI-based air quality PM2.5 forecasting models for developing countries: A case study of Ho Chi Minh City, Vietnam , journal =

Rajnish Rakholia and Quan Le and Khue Vu and Bang Quoc Ho and Ricardo Simon Carbajo , keywords =. AI-based air quality PM2.5 forecasting models for developing countries: A case study of Ho Chi Minh City, Vietnam , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.uclim.2...

2022
[32]

Zeng, Ailing and Chen, Muxi and Zhang, Lei and Xu, Qiang , title =. Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances i...

2023 doi
[33]

The Eleventh International Conference on Learning Representations , year=

A Time Series is Worth 64 Words: Long-term Forecasting with Transformers , author=. The Eleventh International Conference on Learning Representations , year=
[34]

Forecasting: Principles and Practice

Hyndman, \ Robin John\ and George Athanasopoulos. Forecasting: Principles and Practice. 2018

2018
[35]

2008 , publisher=

Forecasting with Exponential Smoothing: The State Space Approach , author=. 2008 , publisher=

2008
[36]

arXiv preprint arXiv:2003.06505 , year=

AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data , author=. arXiv preprint arXiv:2003.06505 , year=

2003 arXiv
[37]

Shchur, Oleksandr and Turkmen, Caner and Erickson, Nick and Shen, Huibin and Shirkov, Alexander and Hu, Tony and Wang, Yuyang , booktitle=
[38]

Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =

Ke, Guolin and Meng, Qi and Finley, Thomas and Wang, Taifeng and Chen, Wei and Ma, Weidong and Ye, Qiwei and Liu, Tie-Yan , title =. Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =. 2017 , isbn =

2017
[39]

DeepAR: Probabilistic forecasting with autoregressive recurrent networks , journal =

David Salinas and Valentin Flunkert and Jan Gasthaus and Tim Januschowski , keywords =. DeepAR: Probabilistic forecasting with autoregressive recurrent networks , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.ijforecast.2019.07.001 , url =

2020 doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Analysis of Air Pollution Data in India between 2015 and 2019 , volume=

Sharma, Disha and Mauzerall, Denise , year=. Analysis of Air Pollution Data in India between 2015 and 2019 , volume=. Aerosol and Air Quality Research , publisher=. doi:10.4209/aaqr.210204 , number=

work page doi:10.4209/aaqr.210204 2015

[2] [2]

and Chen, Yue and Arnold, Steve R

Silver, Ben and Reddington, Carly L. and Chen, Yue and Arnold, Steve R. , journal =. A decade of. 2025 , doi =

2025

[3] [3]

NeurIPS 2025 Workshop on Tackling Climate Change with Machine Learning , url=

Using Time Series Foundation Models for Atmospheric CO2 Concentration Forecasting , author=. NeurIPS 2025 Workshop on Tackling Climate Change with Machine Learning , url=

2025

[4] [4]

Continuous Ambient Air Quality Monitoring (

[5] [5]

Carslaw and Karl Ropkins , keywords =

David C. Carslaw and Karl Ropkins , keywords =. openair — An R package for air quality data analysis , journal =. 2012 , issn =. doi:https://doi.org/10.1016/j.envsoft.2011.09.008 , url =

work page doi:10.1016/j.envsoft.2011.09.008 2012

[6] [6]

China Air Quality Historical Data , author =

[7] [7]

Air Quality Download Service , author =

[8] [8]

Wilson, Robin , url =