The reviewed record of science sign in
Pith

arxiv: 2607.06504 · v1 · pith:5CGJ47TD · submitted 2026-07-07 · cs.AI

RMISC: A Large-scale Real-world Multivariate Corpus for Time Series Foundation Models

Qian Sun , Yong-Ming Tian , Jia-Wei Huang , Cheng Feng , Shao-Qun Zhang This is my paper

Reviewed by Pith2026-07-08 03:23 UTCglm-5.2pith:5CGJ47TDopen to challenge →

classification cs.AI
keywords multivariatereal-worldtsfmsdatatimeseriescorpusgeneralization
0
0 comments X

The pith

Real-world multivariate data beats synthetic for time-series models

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

The paper builds RMISC, a corpus of roughly 200 real-world multivariate time-series datasets totaling 142 billion time points across energy, finance, environment, industry, traffic, and other domains. The central question is whether pretraining time-series foundation models on this real multivariate data yields better zero-shot forecasting than pretraining on synthetic multivariate data, which is what current leading models primarily use. The authors pretrain four representative TSFMs (Chronos-2, GTT, Moirai-2.0, TimesFM-2.5) on seven combinations of three corpus types — real-world univariate (RU), synthetic multivariate (SM), and real-world multivariate (RM) — and evaluate on standard out-of-distribution benchmarks. The key finding is that incorporating the real-world multivariate corpus consistently improves generalization: the RM corpus outranks the SM corpus on average, and the three-way combination RU+SM+RM achieves the best average rank, reducing average MASE by 4.476% relative to the widely used RU+SM combination. The advantage is most pronounced on tasks with complex cross-variable dependencies, where target variables and covariates do not follow simple shared trends.

Core claim

Real-world multivariate time-series data provides cross-variable dependencies and temporal dynamics that synthetic generators cannot replicate, and pretraining on a combination of real-world univariate, synthetic multivariate, and real-world multivariate data produces the strongest zero-shot forecasting models. The paper demonstrates this by showing that the three-source corpus (RU+SM+RM) achieves the best average rank across four models and two benchmarks, and by providing a case study where the model trained on real multivariate data achieves a MASE of 0.2134 versus 0.7697 for the synthetic-trained model on a task with complex covariate relationships.

What carries the argument

The RMISC corpus itself — a curated, schema-unified collection of approximately 200 real-world multivariate datasets with explicit target-covariate annotations, spanning diverse domains and sampling frequencies. The experimental design tests all seven non-empty subsets of three data sources (RU, SM, RM) across four TSFMs, using average ranking across benchmarks to isolate the effect of training data composition on generalization.

If this is right

  • TSFM developers should incorporate real-world multivariate data into pretraining corpora rather than relying on synthetic data alone, particularly when downstream tasks involve complex covariate relationships.
  • Synthetic data generators for multivariate time series may need to improve their ability to produce realistic cross-variable dependencies, since the gap between synthetic and real data is most visible on tasks with non-trivial inter-variable dynamics.
  • The RMISC corpus can serve as a standardized pretraining and evaluation resource for the TSFM community, enabling reproducible comparisons of data composition strategies.
  • The finding that a balanced three-source combination outperforms any single source suggests that data diversity matters more than data purity for zero-shot generalization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the advantage of real multivariate data stems specifically from complex cross-variable dependencies, then the benefit should scale with the number and richness of covariates in downstream tasks — a prediction that could be tested by stratifying benchmark performance by covariate dimensionality.
  • The short pretraining (1-2 epochs on 20M instances) may mean the results reflect early-training dynamics rather than converged behavior; longer training could either widen or narrow the gap between SM and RM corpora if models eventually learn cross-variable patterns from synthetic data given more compute.
  • Domain balance in the RMISC corpus may be a confound: if the balanced subset is used, the improvement could partly come from domain coverage rather than the real-vs-synthetic distinction per se.

Load-bearing premise

The comparison between synthetic multivariate and real-world multivariate data depends on the authors' reproduction of the synthetic data pipeline being faithful to the original, since the exact synthetic corpus used in prior work is not publicly released. If their reproduction produces lower-quality synthetic data than the original, the apparent advantage of real-world data could be inflated.

What would settle it

If a model pretrained on a higher-quality or larger synthetic multivariate corpus matched or exceeded the performance of the RM corpus on the same benchmarks, the claim that real-world multivariate data is intrinsically superior would be weakened.

Figures

Figures reproduced from arXiv: 2607.06504 by Cheng Feng, Jia-Wei Huang, Qian Sun, Shao-Qun Zhang, Yong-Ming Tian.

Figure 1
Figure 1. Figure 1: The modeling workflow of univariate and multivariate time series foundation models on corpora. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall construction pipeline of the RMISC corpus. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training and ID loss curves of four TSFMs on different training corpora of the first epoch. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: ID loss curves of Chronos-2 and GTT on different training corpora of the second epoch. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The average ranking related to MASE of seven corpora across all TSFMs and benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MASE results of different training corpora on OOD benchmarks for (a) Chronos-2, (b) GTT, (c) [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Forecasts generated by Chronos-2 models which are pretrained on the SM corpus and the RM cor [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Domain-wise scale statistics of the proposed dataset. [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Length-dimensionality landscape of all subdatasets. [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Sampling frequency distribution across domains. [PITH_FULL_IMAGE:figures/full_fig_p031_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Data quality distribution across domains. [PITH_FULL_IMAGE:figures/full_fig_p031_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The changes in benchmark MASE scores from the first to the second epoch of different training [PITH_FULL_IMAGE:figures/full_fig_p033_12.png] view at source ↗
read the original abstract

Recent years have witnessed the emergence of multivariate modeling using time series foundation models (TSFMs), which achieve advanced zero-shot generalization. Modern multivariate TSFMs are predominantly pretrained on multivariate synthetic data, which is easier to scale but may fail to capture the complex temporal dynamics and cross-variable relationships present in real-world time series. This raises a key question: Whether and to what extent the leading TSFMs trained with the real-world corpus perform better than those trained with synthetic data? To answer this, we establish the RMISC corpus, a considerably large-scale, high-quality, openly accessible, real-world, and multivariate time series archive that contains around 200 datasets and 142 billion time points across diverse domains. Furthermore, we pretrain four advanced TSFMs on univariate, synthetic multivariate, and real-world multivariate data and evaluate their zero-shot generalization capabilities on standard in-distribution and out-of-distribution benchmarks. Experimental results show that incorporating real-world multivariate data predominantly improves the generalization performance for both univariate and multivariate TSFMs. These results provide a deeper understanding of how real-world multivariate data contributes to the development of stronger TSFMs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 6 minor

Summary. The paper introduces RMISC, a large-scale real-world multivariate time series corpus comprising approximately 200 datasets and 142 billion time points across diverse domains. The authors use this corpus to systematically investigate whether real-world multivariate data improves pretraining of time series foundation models (TSFMs) compared to synthetic multivariate data. Four TSFMs (Chronos-2, GTT, Moirai-2.0, TimesFM-2.5) are pretrained on seven combinations of real-world univariate (RU), synthetic multivariate (SM), and real-world multivariate (RM) corpora, then evaluated on in-distribution validation sets and two out-of-distribution benchmarks (GIFT-Eval, fev-bench) with data-leakage exclusion. The central claim is that the RU+SM+RM combination achieves the best average rank, yielding a 4.476% average MASE reduction over the widely used RU+SM corpus.

Significance. The RMISC corpus itself is a genuine contribution: it is large-scale, openly accessible via Hugging Face, well-organized with standardized metadata, and covers diverse domains and frequencies. The experimental design—four models, seven corpora, two OOD benchmarks with leakage exclusion—is reasonable and addresses a timely question for the TSFM community. The case studies in §3.4 provide useful qualitative evidence for the value of real-world multivariate data in learning complex cross-variable dependencies. However, the quantitative claims rest on single-run experiments without variance estimates, which limits the strength of the conclusions that can be drawn.

major comments (3)
  1. §3.3: The paper states that 'the RM corpus consistently outranks the SM corpus under both single-source-corpus and two-source-corpus pretraining.' This claim is contradicted by Table 4. For Chronos-2, SM outperforms RM on fev-multi (1.039 vs 1.082) and GIFT-uni (1.003 vs 1.034). For GTT, SM outperforms RM on fev-uni (3.056 vs 2.598 is actually RM better, but SM is better on GIFT-uni: 1.116 vs 1.094 — wait, RM is better there). Let me re-examine: for GTT, SM outperforms RM on fev-uni (3.056 vs 2.598 — actually RM is better). The authors themselves later acknowledge that 'neither the SM corpus nor the RM corpus is uniformly superior to the other.' The 'consistently outranks' language in the earlier paragraph is misleading and should be corrected to match the more nuanced assessment the authors provide later in the same section.
  2. §3.3 and Table 4: The headline 4.476% average MASE reduction of RU+SM+RM over RU+SM is based on single-run experiments with no confidence intervals, standard errors, or significance tests. With 70–120M parameter models trained for only 1–2 epochs on 20M sampled instances (§3.1, Table 2), run-to-run variance from initialization and data sampling could plausibly produce MASE shifts of comparable magnitude. Two of the 16 model×benchmark cells actually show degradation (Chronos-2 GIFT-uni: 0.973 vs 0.937; Moirai-2.0 GIFT-uni: 0.979 vs 0.972). While the directional consistency across 14 of 16 cells is suggestive, the authors should either (a) run multiple seeds and report variance/significance, or (b) substantially soften the precision of the 4.476% figure and frame it as a directional trend rather than a measured effect.
  3. §3.2: The convergence argument used to justify 1–2 epochs of pretraining is based on ID loss stabilization. However, ID loss plateauing does not guarantee that the models have fully exploited differences in training data quality. The authors note in Appendix B.2 that the second epoch sometimes degrades OOD performance, which could indicate instability or overfitting rather than true convergence. This is load-bearing because the entire experimental comparison rests on the assumption that 1–2 epochs is sufficient for models to exhibit meaningful differences attributable to training data composition. The authors should discuss this limitation more explicitly, or provide evidence (e.g., learning curves over more epochs) that the relative ordering of corpora stabilizes.
minor comments (6)
  1. Table 1: Several dataset entries show 'Obs.' values without units in the main table (e.g., '1056.50 M' vs '1056.50'). The inconsistency in unit suffixes should be standardized.
  2. §3.1: The SM corpus reproduction is described briefly. Since the exact Chronos-2 synthetic corpus is not publicly released, the authors should provide more details on their reproduction (e.g., specific parameter settings for AR/ETS models, KernelSynth configuration) to allow readers to assess fidelity. This affects the SM-vs-RM comparison, though not the headline RU+SM+RM vs RU+SM comparison.
  3. Figure 5: The average rank figure would benefit from showing individual model results or at least the variance of ranks across models, as the aggregate rank may obscure model-specific patterns visible in Table 4.
  4. Table 4(a), SM+RM row, GIFT-multi WQL column: the value '3.000' appears anomalous compared to surrounding values (~0.3). This may be a typo (perhaps 0.300).
  5. §2, Stage 2: The data processing steps (missing value imputation, outlier handling) are described at a high level. More detail on the specific methods used would improve reproducibility.
  6. The paper uses 'consistently' in several places (abstract, §3.3, §4) to describe improvements from real-world multivariate data. Given the mixed results in Table 4, this language should be tempered throughout.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for a careful and constructive review. The referee raises three major points: (1) the 'consistently outranks' language in §3.3 is contradicted by Table 4 and by our own later nuanced assessment; (2) the headline 4.476% MASE reduction rests on single-run experiments without variance estimates or significance tests; and (3) the convergence argument based on 1–2 epochs of pretraining may be insufficient to guarantee that corpora differences are fully expressed. We agree with all three points and will revise the manuscript accordingly. On point (1), we will correct the language. On point (2), we will run additional seeds and report variance, while also softening the precision of the headline figure. On point (3), we will add extended discussion of the convergence limitation and provide additional learning curve evidence. We cannot fully resolve the concern that 1–2 epochs may be insufficient to express all data-composition differences, as this would require substantially more compute than is available; we acknowledge this as a limitation.

read point-by-point responses

  1. Referee: §3.3: The paper states that 'the RM corpus consistently outranks the SM corpus under both single-source-corpus and two-source-corpus pretraining.' This claim is contradicted by Table 4. For Chronos-2, SM outperforms RM on fev-multi (1.039 vs 1.082) and GIFT-uni (1.003 vs 1.034). The authors themselves later acknowledge that 'neither the SM corpus nor the RM corpus is uniformly superior to the other.' The 'consistently outranks' language in the earlier paragraph is misleading and should be corrected to match the more nuanced assessment the authors provide later in the same section.

    Authors: The referee is correct. The word 'consistently' in the earlier paragraph of §3.3 is too strong and is contradicted by specific cells in Table 4, as well as by our own more nuanced statement later in the same section. For example, for Chronos-2, SM outperforms RM on fev-multi (1.039 vs. 1.082) and GIFT-uni (1.003 vs. 1.034), and for GTT, SM outperforms RM on fev-uni (3.056 vs. 2.598 is actually RM better, but SM is better on GIFT-multi: 1.923 vs. 1.953). We will replace 'consistently outranks' with language that accurately reflects the aggregate-ranking observation—namely that RM achieves a better average rank across all model×benchmark cells, but that the per-cell comparison is mixed—and will ensure the earlier and later paragraphs are consistent. revision: yes

  2. Referee: §3.3 and Table 4: The headline 4.476% average MASE reduction of RU+SM+RM over RU+SM is based on single-run experiments with no confidence intervals, standard errors, or significance tests. With 70–120M parameter models trained for only 1–2 epochs on 20M sampled instances, run-to-run variance from initialization and data sampling could plausibly produce MASE shifts of comparable magnitude. Two of the 16 model×benchmark cells actually show degradation. While the directional consistency across 14 of 16 cells is suggestive, the authors should either (a) run multiple seeds and report variance/significance, or (b) substantially soften the precision of the 4.476% figure and frame it as a directional trend rather than a measured effect.

    Authors: We agree that single-run results without variance estimates limit the strength of the quantitative conclusions. We will take both actions the referee suggests. First, we will run additional random seeds (at least 3 per corpus for each model) and report standard deviations and pairwise significance tests (Wilcoxon signed-rank) for the RU+SM+RM vs. RU+SM comparison. Second, regardless of the outcome of those additional runs, we will reframe the 4.476% figure as an observed directional trend rather than a precise measured effect, and will explicitly note the two cells showing degradation (Chronos-2 GIFT-uni: 0.973 vs. 0.937; Moirai-2.0 GIFT-uni: 0.979 vs. 0.972) in the main text rather than only in the table. We acknowledge that if the multi-seed variance turns out to be large relative to the observed differences, the directional trend itself may need further softening, and we will adjust accordingly based on the additional results. revision: yes

  3. Referee: §3.2: The convergence argument used to justify 1–2 epochs of pretraining is based on ID loss stabilization. However, ID loss plateauing does not guarantee that the models have fully exploited differences in training data quality. The authors note in Appendix B.2 that the second epoch sometimes degrades OOD performance, which could indicate instability or overfitting rather than true convergence. This is load-bearing because the entire experimental comparison rests on the assumption that 1–2 epochs is sufficient for models to exhibit meaningful differences attributable to training data composition. The authors should discuss this limitation more explicitly, or provide evidence (e.g., learning curves over more epochs) that the relative ordering of corpora stabilizes.

    Authors: We agree that ID loss stabilization does not guarantee that all data-composition differences have been fully expressed, and that the second-epoch OOD degradation observed in Appendix B.2 could reflect instability or early overfitting rather than true convergence. We will add an explicit discussion of this limitation in §3.2, acknowledging that 1–2 epochs may be insufficient for the models to fully exploit differences in training data quality and that the relative ordering of corpora could potentially change with longer training. We will also provide extended learning curves (training loss, ID validation loss, and OOD MASE) over additional epochs for at least one representative model and corpus pair, to give partial evidence about whether the relative ordering stabilizes. However, we cannot fully resolve this concern within our compute budget: running all four models across all seven corpora for many more epochs is not feasible. We will state this as a standing limitation of the experimental design. revision: partial

standing simulated objections not resolved
  • We cannot fully rule out the possibility that longer pretraining (more than 2 epochs) would change the relative ordering of corpora, as the referee's third comment suggests. Running all 4 models × 7 corpora for substantially more epochs is beyond our available compute budget. We will acknowledge this as a limitation and provide partial evidence from extended runs on a subset, but a complete resolution is not possible at this time.

Circularity Check

0 steps flagged

No significant circularity found; the paper is an empirical corpus-building study with externally sourced data and models

full rationale

This paper builds a real-world multivariate time series corpus (RMISC) from ~200 publicly available external datasets and uses four existing TSFM architectures (Chronos-2, GTT, Moirai-2.0, TimesFM-2.5) to compare pretraining on different data sources. The central claim—that RU+SM+RM achieves 4.476% average MASE reduction over RU+SM—is an empirical experimental result, not a derivation that reduces to its inputs by construction. The corpus is assembled from external sources, the models are existing architectures, and the comparison is between different training data compositions. No step in the paper's argument chain involves defining a quantity in terms of what it claims to predict, fitting a parameter to a subset and calling it a prediction on closely related data, or invoking a self-citation as the sole justification for a load-bearing premise. The synthetic multivariate corpus is reproduced from Chronos-2's published pipeline (external citation [16]), not from the authors' own prior work. One author (Cheng Feng) co-authored GTT [24], which is used as one of four model architectures, but this is a tool usage, not a circular derivation step. The absence of significance testing and limited training scale are legitimate correctness/statistical concerns, but they are not circularity: the experimental results are not forced by construction. The paper is self-contained against external benchmarks (GIFT-Eval, fev-bench) with data leakage controls, and the findings, while possibly noisy, are not tautological.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new entities, particles, forces, or postulated objects are introduced. The corpus is an aggregation of existing datasets.

free parameters (4)
  • 20M training instances per corpus
    Fixed sample size for pretraining; chosen by the authors, not derived from theory. Affects all comparisons equally but limits scale.
  • Context length range 64-1984
    Chosen pretraining window range; affects what patterns models can learn.
  • Maximum 24 channels per sample
    Caps multivariate dimensionality during training; chosen by authors, affects what cross-variable dependencies can be learned.
  • 80/20 train-validation split ratio
    Standard but arbitrary; affects ID evaluation.
axioms (4)
  • domain assumption The reproduced synthetic multivariate (SM) corpus faithfully represents the Chronos-2 synthetic data pipeline
    Stated in §3.1: 'we reproduce this pipeline to construct our own synthetic multivariate time series dataset.' The entire SM-vs-RM comparison depends on this being a fair reproduction.
  • ad hoc to paper 1-2 epochs of pretraining is sufficient for models to exhibit meaningful differences attributable to training data quality
    The paper concludes convergence after 1-2 epochs (§3.2) but does not verify that longer training would not change the relative rankings of corpora.
  • domain assumption GIFT-Eval and fev-bench are representative OOD benchmarks for evaluating TSFM generalization
    Standard benchmarks used in the field; reasonable but limits conclusions to these specific evaluation sets.
  • standard math MASE and WQL are appropriate metrics for cross-corpus comparison
    Standard forecasting metrics; widely accepted in the literature.

pith-pipeline@v1.1.0-glm · 52019 in / 3525 out tokens · 149879 ms · 2026-07-08T03:23:44.730816+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

145 extracted references · 145 canonical work pages · 17 internal anchors

  1. [1]

    Foundation models for time series analysis: A tutorial and survey

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. Foundation models for time series analysis: A tutorial and survey. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6555–6565, 2024

    work page 2024

  2. [2]

    Benchmarking foundation models for time-series forecasting: Zero-shot, few-shot, and full-shot evaluations.Computer Sciences & Mathematics Forum, 11(1):32, 2025

    Fr ´ed´eric Montet, Benjamin Pasquier, and Beat Wolf. Benchmarking foundation models for time-series forecasting: Zero-shot, few-shot, and full-shot evaluations.Computer Sciences & Mathematics Forum, 11(1):32, 2025

    work page 2025

  3. [3]

    OTexts, 2018

    Rob J Hyndman and George Athanasopoulos.Forecasting: Principles and Practice. OTexts, 2018

    work page 2018

  4. [4]

    Some recent advances in forecasting and control.Journal of the Royal Statistical Society, 17(2):91–109, 1968

    George EP Box and Gwilym M Jenkins. Some recent advances in forecasting and control.Journal of the Royal Statistical Society, 17(2):91–109, 1968. 13

    work page 1968

  5. [5]

    Long short-term memory.Neural Computation, 9(8):1735– 1780, 1997

    Sepp Hochreiter and J ¨urgen Schmidhuber. Long short-term memory.Neural Computation, 9(8):1735– 1780, 1997

    work page 1997

  6. [6]

    Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling.arXiv preprint arXiv:1412.3555, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  7. [7]

    Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting

    Rajat Sen, Hsiang-Fu Yu, and Inderjit S Dhillon. Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting. InAdvances in Neural Information Processing Systems 32, 2019

    work page 2019

  8. [8]

    NHITS: Neural hierarchical interpolation for time series forecasting

    Cristian Challu, Kin G Olivares, Boris N Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, and Artur Dubrawski. NHITS: Neural hierarchical interpolation for time series forecasting. InProceed- ings of the 37th AAAI Conference on Artificial Intelligence, pages 6989–6997, 2023

    work page 2023

  9. [9]

    Temporal fusion transformers for inter- pretable multi-horizon time series forecasting.International Journal of Forecasting, 37(4):1748–1764, 2021

    Bryan Lim, Sercan ¨O Arık, Nicolas Loeff, and Tomas Pfister. Temporal fusion transformers for inter- pretable multi-horizon time series forecasting.International Journal of Forecasting, 37(4):1748–1764, 2021

    work page 2021

  10. [10]

    Predictive maintenance in Industry 4.0: A survey of planning models and machine learning techniques.PeerJ Computer Science, 10:e2016, 2024

    Ida Hector and Rukmani Panjanathan. Predictive maintenance in Industry 4.0: A survey of planning models and machine learning techniques.PeerJ Computer Science, 10:e2016, 2024

    work page 2024

  11. [11]

    Financial time series forecast- ing with deep learning: A systematic literature review: 2005–2019.Applied Soft Computing, 90:106181, 2020

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. Financial time series forecast- ing with deep learning: A systematic literature review: 2005–2019.Applied Soft Computing, 90:106181, 2020

    work page 2005

  12. [12]

    Time series prediction using deep learning methods in healthcare.ACM Transactions on Management Information Systems, 14(1):1–29, 2023

    Mohammad Amin Morid, Olivia R Liu Sheng, and Joseph Dunbar. Time series prediction using deep learning methods in healthcare.ACM Transactions on Management Information Systems, 14(1):1–29, 2023

    work page 2023

  13. [13]

    Springer, 2010

    Manfred Mudelsee.Climate Time Series Analysis: Classical Statistical and Bootstrap Methods. Springer, 2010

    work page 2010

  14. [14]

    Probabilistic electric load forecasting: A tutorial review.International Journal of Forecasting, 32(3):914–938, 2016

    Tao Hong and Shu Fan. Probabilistic electric load forecasting: A tutorial review.International Journal of Forecasting, 32(3):914–938, 2016

    work page 2016

  15. [15]

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. InProceedings of the 6th International Conference on Learning Repre- sentations, 2018

    work page 2018

  16. [16]

    Chronos-2: From Univariate to Universal Forecasting

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris K ¨uken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, et al. Chronos-2: From uni- variate to universal forecasting.arXiv preprint arXiv:2510.15821, 2025. 14

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Timer-XL: Long-context transformers for unified time series forecasting

    Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Timer-XL: Long-context transformers for unified time series forecasting. InProceedings of the 13th International Conference on Learning Representations, 2025

    work page 2025

  18. [18]

    Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting

    Yunhao Zhang and Junchi Yan. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. InProceedings of the 11th International Conference on Learning Representations, 2023

    work page 2023

  19. [19]

    itrans- former: Inverted transformers are effective for time series forecasting

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itrans- former: Inverted transformers are effective for time series forecasting. InProceedings of the 12th Inter- national Conference on Learning Representations, 2024

    work page 2024

  20. [20]

    Empowering Time Series Analysis with Synthetic Data: A Survey and Outlook in the Era of Foundation Models

    Xu Liu, Taha Aksu, Juncheng Liu, Qingsong Wen, Yuxuan Liang, Caiming Xiong, Silvio Savarese, Doyen Sahoo, Junnan Li, and Chenghao Liu. Empowering time series analysis with synthetic data: A survey and outlook in the era of foundation models.arXiv preprint arXiv:2503.11411, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    Uncovering zero- shot generalization gaps in time-series foundation models using real-world videos.arXiv preprint arXiv:2509.26347, 2025

    Lujun Li, Lama Sleem, Yiqun Wang, Yangjie Xu, Niccol `o Gentile, and Radu State. Uncovering zero- shot generalization gaps in time-series foundation models using real-world videos.arXiv preprint arXiv:2509.26347, 2025

    work page arXiv 2025

  22. [22]

    Zero-Shot Time Series Forecasting with Covariates via In-Context Learning

    Andreas Auer, Raghul Parthipan, Pedro Mercado, Abdul Fatir Ansari, Lorenzo Stella, Bernie Wang, Michael Bohlke-Schneider, and Syama Sundar Rangapuram. Zero-shot time series forecasting with co- variates via in-context learning.arXiv preprint arXiv:2506.03128, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    This time is different: An observability perspective on time series foundation models

    Ben Cohen, Emaad Khwaja, Youssef Doubli, Salahidine Lemaachi, Chris Lettieri, Charles Masson, Hugo Miccinilli, Elise Ram ´e, Qiqi Ren, Afshin Rostamizadeh, et al. This time is different: An observability perspective on time series foundation models. InAdvances in Neural Information Processing Systems 38, pages 50907–50951, 2026

    work page 2026

  24. [24]

    Cheng Feng, Long Huang, and Denis Krompass. Only the curve shape matters: Training foundation models for zero-shot multivariate time series forecasting through next curve shape prediction.arXiv preprint arXiv:2402.07570, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    From tables to time: Extending TabPFN- v2 to time series forecasting.arXiv preprint arXiv:2501.02945, 2025

    Shi Bin Hoo, Samuel M ¨uller, David Salinas, and Frank Hutter. From tables to time: Extending TabPFN- v2 to time series forecasting.arXiv preprint arXiv:2501.02945, 2025

    work page arXiv 2025

  26. [26]

    Uni- fied training of universal time series forecasting transformers

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Uni- fied training of universal time series forecasting transformers. InProceedings of the 41st International Conference on Machine Learning, 2024

    work page 2024

  27. [27]

    Generative adversarial networks in time series: A systematic literature review.ACM Computing Surveys, 55(10):1–31, 2023

    Eoin Brophy, Zhengwei Wang, Qi She, and Tom´as Ward. Generative adversarial networks in time series: A systematic literature review.ACM Computing Surveys, 55(10):1–31, 2023. 15

    work page 2023

  28. [28]

    TimeVAE: A Variational Auto-Encoder for Multivariate Time Series Generation

    Abhyuday Desai, Cynthia Freeman, Zuhui Wang, and Ian Beaver. TimeV AE: A variational auto-encoder for multivariate time series generation.arXiv preprint arXiv:2111.08095, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  29. [29]

    Diffusion-TS: Interpretable Diffusion for General Time Series Generation

    Xinyu Yuan and Yan Qiao. Diffusion-TS: Interpretable diffusion for general time series generation.arXiv preprint arXiv:2403.01742, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    Chronos: Learning the Language of Time Series

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Olek- sandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, et al. Chronos: Learning the language of time series.arXiv preprint arXiv:2403.07815, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [31]

    John Wiley & Sons, 2015

    George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung.Time Series Analysis: Forecasting and Control. John Wiley & Sons, 2015

    work page 2015

  32. [32]

    Hyndman, Anne B

    Rob J. Hyndman, Anne B. Koehler, J. Keith Ord, and Ralph D. Snyder.Forecasting With Exponential Smoothing: The State Space Approach. Springer, 2008

    work page 2008

  33. [33]

    A methodology for validating diversity in synthetic time series generation.MethodsX, 8:101459, 2021

    Fouad Bahrpeyma, Mark Roantree, Paolo Cappellari, Michael Scriney, and Andrew McCarren. A methodology for validating diversity in synthetic time series generation.MethodsX, 8:101459, 2021

    work page 2021

  34. [34]

    fev-bench: A Realistic Benchmark for Time Series Forecasting

    Oleksandr Shchur, Abdul Fatir Ansari, Caner Turkmen, Lorenzo Stella, Nick Erickson, Pablo Guerron, Michael Bohlke-Schneider, and Yuyang Wang. Fev-bench: A realistic benchmark for time series fore- casting.arXiv preprint arXiv:2509.26468, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  35. [35]

    GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation

    Taha Aksu, Gerald Woo, Juncheng Liu, Xu Liu, Chenghao Liu, Silvio Savarese, Caiming Xiong, and Doyen Sahoo. GIFT-eval: A benchmark for general time series forecasting model evaluation.arXiv preprint arXiv:2410.10393, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Monash Time Series Forecasting Archive

    Rakshitha Godahewa, Christoph Bergmeir, Geoffrey I Webb, Rob J Hyndman, and Pablo Montero- Manso. Monash time series forecasting archive.arXiv preprint arXiv:2105.06643, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  37. [37]

    MO- MENT: A family of open time-series foundation models

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. MO- MENT: A family of open time-series foundation models. InProceedings of the 41st International Con- ference on Machine Learning, pages 16115–16152, 2024

    work page 2024

  38. [38]

    A decoder-only foundation model for time-series forecasting

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. InProceedings of the 41st International Conference on Machine Learning, pages 10148–10167, 2024

    work page 2024

  39. [39]

    Appliance consumption signature database and recognition test protocols

    Christophe Gisler, Philippe Bontron, Omar Abou Khaled, and Jean Hennebert. Appliance consumption signature database and recognition test protocols. InProceedings of the 8th International Workshop on Systems, Signal Processing and their Applications, pages 258–263, 2013

    work page 2013

  40. [40]

    Fast and accurate time series classification with WEASEL

    Patrick Schafer and Ulf Leser. Fast and accurate time series classification with WEASEL. InProceedings of the 2017 ACM Conference on Information and Knowledge Management, pages 637–646, 2017. 16

    work page 2017

  41. [41]

    Time-MoE: Billion-scale time series foundation models with mixture of experts

    Shi Xiaoming, Wang Shiyu, Nie Yuqi, Li Dianqi, Ye Zhou, Wen Qingsong, and Ming Jin. Time-MoE: Billion-scale time series foundation models with mixture of experts. InProceedings of the 13th Interna- tional Conference on Learning Representations, 2025

    work page 2025

  42. [42]

    Meehl, Catherine A

    Veronika Eyring, Sandrine Bony, Gerald A. Meehl, Catherine A. Senior, Bjorn Stevens, Ronald J. Stouf- fer, and Karl E. Taylor. Overview of the coupled model intercomparison project phase 6 (CMIP6) exper- imental design and organization.Geoscientific Model Development, 9:1937–1958, 2016

    work page 1937

  43. [43]

    Improving S&P stock prediction with time series stock similarity

    Lior Sidi. Improving S&P stock prediction with time series stock similarity.arXiv preprint arXiv:2002.05784, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2002

  44. [44]

    CSTS: A Benchmark for the Discovery of Correlation Structures in Time Series Clustering

    Isabella Degen, Zahraa S Abdallah, Henry W J Reeve, and Kate Robson Brown. CSTS: A benchmark for the discovery of correlation structures in time series clustering.arXiv preprint arXiv:2505.14596, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  45. [45]

    Appliances energy prediction.UCI Machine Learning Repository, 2017

    Luis Candanedo. Appliances energy prediction.UCI Machine Learning Repository, 2017

    work page 2017

  46. [46]

    Aus- tralian electricity demand dataset.Zenodo, 2021

    Rakshitha Godahewa, Christoph Bergmeir, Geoff Webb, Rob Hyndman, and Pablo Montero-Manso. Aus- tralian electricity demand dataset.Zenodo, 2021

    work page 2021

  47. [47]

    Resource central: Understanding and predicting workloads for improved resource management in large cloud platforms

    Eli Cortez, Anand Bonde, Alexandre Muzio, Mark Russinovich, Marcus Fontoura, and Ricardo Bian- chini. Resource central: Understanding and predicting workloads for improved resource management in large cloud platforms. InProceedings of the 26th Symposium on Operating Systems Principles, pages 153–167, 2017

    work page 2017

  48. [48]

    Shape classifier based on generalized probabilistic descent method with hidden Markov descriptor

    Ninad Thakoor and Jean Gao. Shape classifier based on generalized probabilistic descent method with hidden Markov descriptor. InProceedings of the 10th IEEE International Conference on Computer Vision, 2005

    work page 2005

  49. [49]

    The building data genome project 2, energy meter data from the ASHRAE great energy predictor III competition.Scientific Data, 7(368), 2020

    Clayton Miller, Archan Kathirgamanathan, Bruno Picchetti, et al. The building data genome project 2, energy meter data from the ASHRAE great energy predictor III competition.Scientific Data, 7(368), 2020

    work page 2020

  50. [50]

    ClimateLearn: Benchmarking machine learning for weather and climate modeling

    Tung Nguyen, Jason Kyle Jewik, Hritik Bansal, Prakhar Sharma, and Aditya Grover. ClimateLearn: Benchmarking machine learning for weather and climate modeling. InAdvances in Neural Information Processing Systems 36, pages 75009–75025, 2023

    work page 2023

  51. [51]

    Hasell, E

    J. Hasell, E. Mathieu, D. Beltekian, et al. A cross-country database of COVID-19 testing.Scientific Data, 7:345, 2020

    work page 2020

  52. [52]

    Mathieu, H

    E. Mathieu, H. Ritchie, E. Ortiz-Ospina, et al. A global database of COVID-19 vaccinations.Nature Human Behaviour, 2021

    work page 2021

  53. [53]

    Gas sensor array temperature modulation.UCI Machine Learning Repository, 2018

    Javier Burgus. Gas sensor array temperature modulation.UCI Machine Learning Repository, 2018. 17

    work page 2018

  54. [54]

    COVID-19 deaths dataset.Zenodo, 2020

    Rakshitha Godahewa, Christoph Bergmeir, Geoff Webb, Rob Hyndman, and Pablo Montero-Manso. COVID-19 deaths dataset.Zenodo, 2020

    work page 2020

  55. [55]

    COVID-19 mobility dataset (with missing values).Zenodo, 2021

    Rakshitha Godahewa, Christoph Bergmeir, Geoff Webb, Rob Hyndman, and Pablo Montero-Manso. COVID-19 mobility dataset (with missing values).Zenodo, 2021

    work page 2021

  56. [56]

    KDD cup dataset (with missing values).Zenodo, 2020

    Rakshitha Godahewa, Christoph Bergmeir, Geoff Webb, Rob Hyndman, and Pablo Montero-Manso. KDD cup dataset (with missing values).Zenodo, 2020

    work page 2020

  57. [57]

    Oikolab weather dataset.Zenodo, 2021

    Rakshitha Godahewa, Christoph Bergmeir, Geoff Webb, Rob Hyndman, and Pablo Montero-Manso. Oikolab weather dataset.Zenodo, 2021

    work page 2021

  58. [58]

    Krilova, I

    N. Krilova, I. Kastalskiy, V . Kazantsev, V . A. Makarov, and S. Lobov. EMG data for gestures.UCI Machine Learning Repository, 2018

    work page 2018

  59. [59]

    PM2.5 data of five Chinese cities.UCI Machine Learning Repository, 2016

    Song Chen. PM2.5 data of five Chinese cities.UCI Machine Learning Repository, 2016

    work page 2016

  60. [60]

    Abdulaal and T

    A. Abdulaal and T. Lancewicki. Real-time synchronization in neural networks for multivariate time series anomaly detection. InProceedings of the 2021 IEEE International Conference on Acoustics, Speech, and Signal Processing, 2021

    work page 2021

  61. [61]

    Abdulaal, Z

    A. Abdulaal, Z. Liu, and T. Lancewicki. Practical approach to asynchronous multivariate time series anomaly detection and localization. InProceedings of the 27th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2021

    work page 2021

  62. [62]

    BuildingsBench: A large-scale dataset of 900k buildings and benchmark for short-term load forecasting

    Patrick Emami, Abhijeet Sahu, and Peter Graf. BuildingsBench: A large-scale dataset of 900k buildings and benchmark for short-term load forecasting. InAdvances in Neural Information Processing Systems 36, pages 19823–19857, 2023

    work page 2023

  63. [63]

    Sub- seasonalClimateUSA: A dataset for subseasonal forecasting and benchmarking

    Soukayna Mouatadid, Paulo Orenstein, Genevieve Flaspohler, Miruna Oprescu, Judah Cohen, Franklyn Wang, Sean Knight, Maria Geogdzhayeva, Sam Levang, Ernest Fraenkel, and Lester Mackey. Sub- seasonalClimateUSA: A dataset for subseasonal forecasting and benchmarking. InAdvances in Neural Information Processing Systems 36, 2023

    work page 2023

  64. [64]

    Grundy, Alison E

    Evi Yemini, Tomas Jucikas, Luke J. Grundy, Alison E. Brown, and William R. Schafer. A database of Caenorhabditis elegans behavioral phenotypes.Nature Methods, 10(9):877–879, 2013

    work page 2013

  65. [65]

    Temperature rain dataset without missing values.Zenodo, 2021

    Rakshitha Godahewa, Christoph Bergmeir, Geoff Webb, Rob Hyndman, and Pablo Montero-Manso. Temperature rain dataset without missing values.Zenodo, 2021

    work page 2021

  66. [66]

    Yap, Matthew Amos, and Flora D

    Arian Prabowo, Xiachong Lin, Imran Razzak, Hao Xue, Emily W. Yap, Matthew Amos, and Flora D. Salim. BTS: Building timeseries dataset: Empowering large-scale building analytics. InAdvances in Neural Information Processing Systems 38, 2024. 18

    work page 2024

  67. [67]

    Dueben, Sebastian Scher, Jonathan A

    Stephan Rasp, Peter D. Dueben, Sebastian Scher, Jonathan A. Weyn, Soukayna Mouatadid, and Nils Thuerey. WeatherBench: A benchmark data set for data-driven weather forecasting.Journal of Advances in Modeling Earth Systems, 12(11), 2020

    work page 2020

  68. [68]

    BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation

    Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, et al. Behavior-1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation.arXiv preprint arXiv:2403.09227, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  69. [69]

    Multivariate gait data.UCI Machine Learning Reposi- tory, 2016

    Nathaniel Helwig and Elizabeth Hsiao-Wecksler. Multivariate gait data.UCI Machine Learning Reposi- tory, 2016

    work page 2016

  70. [70]

    HAR70+.UCI Machine Learning Repository, 2023

    Aleksej Logacjov and Astrid Ustad. HAR70+.UCI Machine Learning Repository, 2023

    work page 2023

  71. [71]

    Weather dataset.Zenodo, 2020

    Rakshitha Godahewa, Christoph Bergmeir, Geoff Webb, Pablo Montero-Manso, and Rob Hyndman. Weather dataset.Zenodo, 2020

    work page 2020

  72. [72]

    Beaver, Raymond C

    Justin M. Beaver, Raymond C. Borges-Hink, and Mark A. Buckner. An evaluation of machine learning methods to detect malicious SCADA communications. InProceedings of the 12th International Confer- ence on Machine Learning and Applications, pages 54–59, 2013

    work page 2013

  73. [73]

    Aleksej Logacjov, Atle Kongsvold, Kerstin Bach, Hilde Bremseth B˚ardstu, and Paul Jarle Mork. HARTH. UCI Machine Learning Repository, 2023

    work page 2023

  74. [74]

    Gas sensor array under dynamic gas mixtures.UCI Machine Learning Repository, 2015

    Jordi Fonollosa. Gas sensor array under dynamic gas mixtures.UCI Machine Learning Repository, 2015

    work page 2015

  75. [75]

    Heterogeneity activity recognition.UCI Machine Learning Repository, 2015

    Henrik Blunck, Sourav Bhattacharya, Thor Prentow, Mikkel Kjrgaard, and Anind Dey. Heterogeneity activity recognition.UCI Machine Learning Repository, 2015

    work page 2015

  76. [76]

    A three-year building operational performance dataset for informing energy efficiency.Dryad, 2022

    Tianzhen Hong, Na Luo, David Blum, and Zhe Wang. A three-year building operational performance dataset for informing energy efficiency.Dryad, 2022

    work page 2022

  77. [77]

    Hungarian chickenpox cases.UCI Machine Learning Repository, 2021

    UCI. Hungarian chickenpox cases.UCI Machine Learning Repository, 2021

    work page 2021

  78. [78]

    Occupancy detection.UCI Machine Learning Repository, 2016

    Luis Candanedo. Occupancy detection.UCI Machine Learning Repository, 2016

    work page 2016

  79. [79]

    Informer: Beyond efficient Transformer for long sequence time-series forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient Transformer for long sequence time-series forecasting. InProceedings of the 35th AAAI Conference on Artificial Intelligence, pages 11106–11115, 2021

    work page 2021

  80. [80]

    Price graphs: Utilizing the structural information of financial time series for stock prediction.Information Sciences, 588:405–424, 2022

    Junran Wu, Ke Xu, Xueyuan Chen, Shangzhe Li, and Jichang Zhao. Price graphs: Utilizing the structural information of financial time series for stock prediction.Information Sciences, 588:405–424, 2022

    work page 2022

Showing first 80 references.