Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

BLAST: Balanced Sampling Time Series Corpus for Universal Forecasting Models

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Balancing time-series training data by statistical pattern, rather than by domain label, lets universal forecasting models match or beat far larger, costlier pre-training runs.

desk verdict A genuinely new data-balancing pipeline for time-series pretraining, with one clean controlled experiment, but the headline efficiency claim is confounded and a climate-leakage check is needed. read the letter →

arxiv 2505.17871 v2 pith:QNEQUSD6 submitted 2025-05-23 cs.LG

Zezhi Shao , Yujie Li , Fei Wang , Chengqing Yu , Yisong Fu , Tangwen Qian , Bin Xu , Boyu Diao
show 2 more authors
Yongjun Xu Xueqi Cheng
This is my paper
classification cs.LG
keywords large-scaletimeseriesdatasetbalancedsamplinguniversalforecastingzero-shotdatadiversitygridmixuppre-trainingcorpus
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's thesis is that the composition of a pre-training corpus—not just its size—controls how well universal forecasting models generalise, and that balance across patterns is the missing ingredient. To test this, it builds BLAST, a 321-billion-observation corpus, and samples it by 'pattern' rather than by dataset: each series is compressed into seven statistical fingerprints (stationarity, trend, seasonality, volatility, variance drift, memory, anomaly fraction), projected to a 2D plane, and then drawn uniformly from cells of a grid on that plane. Retraining open models on this balanced corpus outperforms their original releases: TimeMoE_base trained on BLAST reaches 0.325 MSE / 0.368 MAE on the ETT/Weather/GlobalTemp suite with 78.64B tokens on 8 A100 GPUs, versus 0.341 / 0.385 for the original run of 419.43B tokens on 128 A100 GPUs. The authors conclude that pattern-balanced sampling is a substitute for raw scale, cutting GPU cost roughly sixteenfold and token count roughly fivefold while improving zero-shot accuracy.

What carries the argument

The load-bearing machinery is the statistical fingerprint plus the grid built from it. Seven metrics—ADF stationarity (Eq. 5), Mann–Kendall trend and strength (Eq. 6), M-STL seasonality count and strength (Eq. 7), relative volatility (Eq. 8), Lagrange-multiplier scedasticity (Eq. 9), Hurst memorability (Eq. 10), and a z-score anomaly fraction (Eq. 11)—are discretised ($B$ bins per metric), one-hot encoded, and concatenated into a fixed vector $h \in \mathbb{R}^{61}$. UMAP projects $h$ to $\mathbb{R}^2$, and a $100 \times 100$ grid on the unit square becomes the strata: cell $m$ is chosen with probability $1/|\mathcal{G}|$, a series inside it with probability $1/|G_m|$, then a window from that series uniformly (Eq. 14). Grid mixup mixes windows from $\mathrm{Uniform}(1,3)$ randomly chosen cells with Dirichlet($\alpha=1.5$) weights (Eq. 15). The paper's central move is claiming that these grid cells—not dataset or domain labels—are the correct sampling units, because labels do not reliably separate patterns.

What would settle it

Permute the mapping between time series and their seven-metric fingerprints before building the UMAP grid, so each cell contains series whose recorded patterns are not their own, then train TimeMoE_base on this permuted-BLAST corpus under the paper's settings. If average MSE on the ETT/Weather/GlobalTemp suite stays near 0.325, the gains come from grid-sampling mechanics alone; if it rises toward the naive-sampling numbers (e.g., ETTh1 MAE at horizon 96 climbs from 0.376 toward 0.393), the statistical fingerprints themselves carry the balancing signal.

Watch

Extended reading notes

Core claim

The central discovery is that how a pre-training corpus is sampled changes what a universal forecasting model learns more than how many tokens it sees. BLAST's recipe: characterise each of roughly 40 million time series by seven cheap statistical metrics; discretise and one-hot encode them into a fixed 61-dimensional vector; project the vectors to a plane with UMAP; cut the plane into a 100x100 grid whose cells act as implicit clusters; then sample a cell uniformly, a series within it uniformly, and a sliding window from that series. A second stage, grid mixup, mixes up to three windows drawn from distinct cells with Dirichlet weights to fill empty regions of pattern space. On the same raw data, this scheme beats naive sampling and dataset-stratified sampling in both convergence speed and error, and models retrained from scratch on it (TimeMoE, MOIRAI, Chronos) beat their officially released counterparts, the largest win being TimeMoE_base at roughly five times fewer tokens and sixteen times fewer GPUs for better average accuracy.

Load-bearing premise

The seven statistical fingerprints are assumed to capture every difference between time series that matters for forecasting; if two series look alike on these metrics yet forecast very differently, grid cells mix unlike patterns and the balancing argument collapses.

Editorial extensions

If this is right

  • On identical raw data, balanced sampling reaches a fixed validation loss in roughly 35% of the training steps required by naive or stratified sampling, so the same model can be trained much more cheaply.
  • Retraining publicly released models from scratch on BLAST improves zero-shot MASE on the filtered GIFT-Eval tasks (TimeMoE_base 0.760 vs 0.888, MOIRAI_base 0.759 vs 0.812, Chronos_base 0.711 vs 0.740), showing the corpus benefit transfers across architectures.
  • Removing either component degrades all four ETT benchmarks relative to balanced sampling, with the no-grid-sampling ablation falling closest to naive-sampling levels, confirming that cell-uniform selection and cell-mixing fill distinct roles.
  • Grid size behaves like an implicit clustering resolution: $M=100$ is best among the tested values, while both $M=10$ and $M=5000$ degrade toward naive sampling, so the optimum tracks the granularity of patterns in the corpus.
  • Because ETT, Weather, and GlobalTemp are excluded from BLAST, the reported gains are true out-of-distribution zero-shot results relative to the corpus.

Reading between the lines

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

  • The fingerprint-plus-grid recipe is transportable: any domain where cheap statistical descriptors separate informative from redundant data (audio, finance, sensor logs) could reuse grid sampling without labels, though the metric set would need re-selection.
  • The paper's stated rule that metrics should not be directly tied to predictability creates a tension: fingerprints that best capture forecastability would also exclude the hard samples balanced sampling is meant to retain; resolving that trade-off is a natural follow-up.
  • A scaling law linking optimal grid count to corpus size and pattern diversity is missing; the 100x100 grid and UMAP hyperparameters (n_neighbors=100, min_dist=0.9) were validated on synthetic uniform data, so the recipe is not yet parameter-free.
  • The BLAST advantage is measured against corpora whose bias it corrects; the most direct stress test is to apply the same balanced sampler to an unfiltered, deliberately skewed corpus and check whether zero-shot performance remains flat across the skewed and de-skewed versions.
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.

Referee Report

4 major / 6 minor

Summary. This paper introduces BLAST, a large-scale pre-training corpus for universal time series forecasting models, containing 321 billion observations assembled from public datasets. The core proposal is a balanced sampling strategy: each time series is characterized by seven statistical metrics (stationarity, trend, seasonality, volatility, scedasticity, memorability, and anomaly fraction), the resulting feature vectors are embedded into two dimensions with UMAP, the plane is partitioned into grids, and training samples are drawn uniformly per grid, with an additional grid-mixup operation. The authors retrain TimeMoE, MOIRAI, and Chronos from scratch on BLAST and report improved zero-shot performance on six benchmarks and on the filtered GIFT-Eval benchmark, along with a large reduction in training tokens and GPU requirements relative to the original TimeMoE. A controlled comparison of sampling strategies on the same raw data is also presented.

Significance. If the claims hold, this is a significant data-centric contribution: it identifies sampling strategy, rather than raw corpus size alone, as a key driver of pre-training efficiency for universal forecasting models. The public release of the corpus-generation code and the processed data is a practical strength, and the controlled ablation in Table 7 (same raw data, same model, balanced versus naive versus stratified sampling) provides genuine evidence for the core mechanism. The GIFT-Eval evaluation after filtering out data present in prior corpora is also a good practice. However, the headline efficiency claim is not a controlled comparison, the grid-size hyperparameter is selected on benchmarks that later appear in the main results, and the possibility of series-level overlap between the climate-dominated BLAST corpus and the Weather/GlobalTemp evaluation sets is not ruled out. These issues must be addressed before the stronger claims can be accepted.

major comments (4)
  1. [§6.2, Table 1] The flagship efficiency claim is not a controlled comparison. TimeMoE_base-BLAST is trained on 78.64B tokens with batch size 192 on 8 GPUs, whereas the original TimeMoE_base numbers come from a run with 419.43B tokens, batch size 1024, and 128 GPUs, and the two runs also use different training corpora. The average-MSE improvement from 0.341 to 0.325 therefore cannot be uniquely attributed to BLAST. Please provide a matched comparison (same token budget, same batch size, same optimizer schedule) between BLAST and Time-300B, or clearly reframe the claim as a resource-efficiency result conditional on BLAST's data. Table 7 is a useful controlled step, but it only compares sampling strategies on ETT, not against the original TimeMoE run.
  2. [§6.4, Figure 6] The grid-size hyperparameter M is selected by evaluating on the four ETT datasets that later appear in the headline Table 5. The paper chooses M=100 after varying M in {10, 50, 100, 500, 1000, 5000} and measuring MAE/MSE on ETTh1, ETTh2, ETTm1, and ETTm2, and the same datasets are then used to report the main improvements. This is test-set selection and can inflate the reported gains. Please select M on a validation split (for example, a time-based holdout from the training data) or report results across all values of M for all benchmarks; at minimum, state clearly that M was chosen using the ETT test sets and quantify the sensitivity of the results to this choice.
  3. [§6.1.2, Appendix A.1] The statement that 'None of these datasets is included in BLAST' only excludes benchmark datasets by name; it does not rule out series-level overlap. BLAST's raw corpus is dominated by CMIP6 (32.5%), ERA5 (30.0%), and WeatherBench (25.7%), while the Weather and GlobalTemp evaluation sets are station or reanalysis series from related sources. Near-duplicate series may therefore be present despite different dataset names. Please perform a series-level deduplication or similarity check (for example, normalized cross-correlation or nearest-neighbor search) between BLAST and all evaluation series, and report the main results with any overlapping series removed.
  4. [§5.2, Eq. (14)] The method assumes that uniform coverage of the seven hand-picked statistical metrics produces balanced coverage of forecasting-relevant patterns. This assumption is central to the balanced-sampling argument but is not directly validated. Please add an analysis showing that grid cells are homogeneous with respect to holdout forecasting error (for example, comparing within-grid versus between-grid variance of a reference forecaster's error), or provide an ablation using additional or different metrics, to demonstrate that the metric space is sufficient. Without such evidence, the generalization of the balancing benefit beyond the tested ETT/weather settings remains an assumption.
minor comments (6)
  1. [§6.2] The sentence 'The only deviation from the original setup was the batch size for TimeMoE' is inaccurate, because the training corpus and token budget also differ from the original TimeMoE run; please rephrase to avoid confusion.
  2. [§6.6] The phrase 'draws unbiased samples via stratified sampling guided by that PDF' is misleading: grid sampling deliberately upweights under-represented pattern regions, so 'balanced' rather than 'unbiased' would be the correct description.
  3. [§6.4, §6.5] There are typographical errors: 'The reaseaons are' should be 'The reasons are' in §6.4, and 'one-how vector' should be 'one-hot vector' in §6.5.
  4. [Figure 5] Please add axis labels and a clear definition of 'validation loss' and the 'Saved 65% Training Cost' annotation; the current figure caption is ambiguous about what is plotted.
  5. [Appendix A.1, A.4] Please clarify the missing-value policy: series with more than 5% NaNs are filtered out, but the remaining NaNs are kept in the released matrices; it would be helpful to state whether the 5% threshold is applied per series and whether NaN padding in the N×L release is distinguishable from observed NaN values.
  6. [§5.4] Please clarify how the 'approximately 40 million raw time series' count relates to the 321B observations and to the 4096-length sampled segments, so readers can reconcile the corpus statistics.

Circularity Check

1 steps flagged · score 4.0 of 10

Grid-size hyperparameter chosen on ETT benchmarks weakens the ETT part of the headline gain, but the method retains independent controlled evidence.

  1. fitted input called prediction [Section 5.5 (Eq. 14) and Section 6.4, Figure 6]
    "We set M = 100. … Furthermore, we vary the grid size in the sampling stage, setting it to [10, 50, 100, 500, 1000, 5000]. We evaluate the models on four ETT datasets and report their averaged predictive performance."

    The grid size M is the key hyperparameter of BLAST's grid sampling (Eq. 14). Section 6.4 chooses it by sweeping [10, 50, 100, 500, 1000, 5000] and evaluating on 'four ETT datasets'; Section 5.5 then fixes 'We set M = 100.' The same four ETT benchmarks reappear as four of the six datasets in Table 5's headline average and in Table 7's balanced-sampling comparison. Thus the reported ETT improvement of balanced sampling over naive/stratified is partially a result of selecting M on the test benchmarks, i.e., a fitted input presented as a prediction. It does not make the whole method circular—GIFT-Eval (43 excluded tasks) and the controlled same-raw-data ablation retain independent content—but the ETT portion of the SOTA claim is partly a selection artifact.

full rationale

The central construction is not a derivation: BLAST computes seven statistical descriptors, UMAP-embeds them, partitions into grids, and samples/mixes grid cells. None of these equations (Eqs. 5-15) is defined in terms of the evaluation targets, and no load-bearing premise rests on a uniqueness theorem or on the authors' own prior work. The two self-citations ([18], [36]) are background references, not evidence for the method. The strongest internal evidence, Table 7, is a controlled same-raw-data comparison, and GIFT-Eval excludes any data present in BLAST. The one genuine circular element is the grid-size choice: M=100 was selected by evaluating on the same four ETT datasets that later support the main efficiency/accuracy claim, so part of the ETT gain is a selection artifact rather than an out-of-sample prediction. The separate concern that BLAST's dominant climate sources (CMIP6/ERA5/WeatherBench) might overlap Weather/GlobalTemp is a leakage/contamination risk, not a definitional circularity, and falls outside this pass.

Assumptions & free parameters 12 free parameters · 5 assumptions · 0 invented entities

The corpus construction introduces many hand-chosen thresholds, bin counts, and sampling hyperparameters. Several (grid size, mixup coefficients) are chosen via experiments on the same benchmarks used for evaluation, which is the main circularity burden. No new physical or theoretical entities are postulated.

free parameters (12)
  • ADF significance threshold = 0.05
    Used in Eq. 5 to binarize stationarity; affects which series are labeled stationary.
  • Trend strength discretization bins = B=20, [b0,bB]=[-1,1]
    Table 4; hand-chosen bin count and range for Mann-Kendall trend strength.
  • Seasonality strength discretization bins = B=10, [b0,bB]=[0,1]
    Table 4; hand-chosen bin count and range for M-STL seasonality strength.
  • Volatility discretization bins = B=6, [b0,bB]=[0,1.2]
    Table 4; hand-chosen bin count and range for volatility.
  • Memorability discretization bins = B=10, [b0,bB]=[0,1]
    Table 4; hand-chosen bin count and range for Hurst exponent.
  • Anomaly discretization bins = B=4, [b0,bB]=[0,0.16]
    Table 4; hand-chosen bin count and range for anomaly fraction.
  • Anomaly z-score threshold = 1.645
    Eq. 11; one-tailed 95% threshold for outlier detection after z-score normalization.
  • Grid size M = 100
    Section 6.4 tunes M on the ETT benchmarks that also appear in final evaluation; directly controls the balancing granularity.
  • Grid mixup K and alpha = K=3, alpha=1.5
    Section 5.5; number of blended grids and Dirichlet concentration, chosen by hand.
  • Max series length and segment count = 4096, 3 segments
    Appendix A.2.2; long series are truncated to 4096 by averaging/voting over three random segments.
  • UMAP n_neighbors and min_dist = 100, 0.9
    Appendix A.3.2; selected via synthetic-data hyperparameter study, affecting global structure preservation.
  • Filter thresholds = length>=512, NaN<=5%
    Section 5.1 and Appendix A.1; hand-chosen cutoffs for excluding short or missing-heavy series.
assumptions (5)
  • domain assumption Seven statistical metrics are a sufficient representation of time series pattern diversity.
    Section 5.2 and A.2.1 state the metrics characterize patterns; if false, grid balancing is misaligned with predictability-relevant diversity.
  • domain assumption UMAP embeddings preserve global structure enough that uniform 2D grid partitions correspond to pattern clusters.
    Section 5.4 and 6.5 argue UMAP captures both local and global structure; the entire grid sampling scheme depends on this preservation.
  • standard math Sliding-window candidates with stride 1 are exchangeable and naive uniform sampling over W is the natural baseline.
    Definitions 2 and Eq. 2; standard counting of windows under a fixed context length.
  • domain assumption Balanced pattern coverage improves zero-shot generalization and training efficiency.
    This is the paper's central hypothesis; the method is built on it and the experiments test it, but it is an assumption rather than a proven theorem.
  • domain assumption The ETT, Weather, GlobalTemp, and GIFT-Eval benchmarks are representative of universal forecasting performance.
    Section 6.1.2 adopts TimeMoE's benchmark suite; results may not generalize to other domains, a limitation acknowledged only implicitly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BLAST: Balanced Sampling Time Series Corpus for Universal Forecasting Models." pith.science (2026). https://pith.science/paper/QNEQUSD6

@misc{pith2026250517871,
  author       = {Pith},
  title        = {Pith review of: BLAST: Balanced Sampling Time Series Corpus for Universal Forecasting Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QNEQUSD6}},
  note         = {Machine review of arXiv:2505.17871}
}
read the original abstract

The advent of universal time series forecasting models has revolutionized zero-shot forecasting across diverse domains, yet the critical role of data diversity in training these models remains underexplored. Existing large-scale time series datasets often suffer from inherent biases and imbalanced distributions, leading to suboptimal model performance and generalization. To address this gap, we introduce BLAST, a novel pre-training corpus designed to enhance data diversity through a balanced sampling strategy. First, BLAST incorporates 321 billion observations from publicly available datasets and employs a comprehensive suite of statistical metrics to characterize time series patterns. Then, to facilitate pattern-oriented sampling, the data is implicitly clustered using grid-based partitioning. Furthermore, by integrating grid sampling and grid mixup techniques, BLAST ensures a balanced and representative coverage of diverse patterns. Experimental results demonstrate that models pre-trained on BLAST achieve state-of-the-art performance with a fraction of the computational resources and training tokens required by existing methods. Our findings highlight the pivotal role of data diversity in improving both training efficiency and model performance for the universal forecasting task.

Figures

Figures reproduced from arXiv: 2505.17871 by the authors.

Figure 2
Figure 2. The uneven distribution of the raw large-scale time [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Pipeline for the balanced sampling: (i) constructing large-scale time series datasets, (ii) utilizing diverse metrics to comprehensively characterize time series, (iii) generating unified feature vectors and performing dimension reduction to visualize data imbalances, and (iv) implementing grid sampling and grid mixup to enhance the diversity of the training data. W is the candidate sample set, and is formally defin… view at source ↗
Figure 4
Figure 4. Distribution of the raw dataset across key metrics. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Comparison of convergence speeds for different [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The impact of grid size in grid sampling. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ARIES: Relation Assessment and Model Recommendation for Deep Time Series Forecasting

    cs.LG 2025-09 conditional novelty 6.0 of 10

    ARIES shows that deep forecasting models have consistent performance preferences tied to time series properties, and uses those preferences to recommend models for new datasets.

Reference graph

Works this paper leans on

49 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    Amro Kamal Mohamed Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S Morcos. [n. d.]. SemDeDup: Data-efficient learning at web-scale through semantic deduplication. In ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation Models

    work page 2023

  2. [2]

    Taha Aksu, Gerald Woo, Juncheng Liu, Xu Liu, Chenghao Liu, Silvio Savarese, Caiming Xiong, and Doyen Sahoo. 2024. GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation.arXiv preprint arXiv:2410.10393 (2024)

    arXiv 2024

  3. [3]

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebas- tian Pineda Arango, Shubham Kapoor, et al. 2024. Chronos: Learning the Lan- guage of Time Series. Transactions on Machine Learning Research (2024)

    work page 2024

  4. [4]

    Kasun Bandara, Rob J Hyndman, and Christoph Bergmeir. 2021. MSTL: A seasonal- trend decomposition algorithm for time series with multiple seasonal patterns. arXiv preprint arXiv:2107.13462 (2021)

    arXiv 2021

  5. [5]

    Tim Bollerslev. 1986. Generalized autoregressive conditional heteroskedasticity. Journal of econometrics 31, 3 (1986), 307–327

    work page 1986

  6. [6]

    Mouxiang Chen, Lefei Shen, Zhuo Li, Xiaoyun Joy Wang, Jianling Sun, and Chenghao Liu. 2024. Visionts: Visual masked autoencoders are free-lunch zero- shot time series forecasters. arXiv preprint arXiv:2408.17253 (2024)

    arXiv 2024

  7. [7]

    Yanping Chen, Eamonn Keogh, Bing Hu, Nurjahan Begum, Anthony Bagnall, Abdullah Mueen, and Gustavo Batista. 2015. The UCR Time Series Classification Archive. www.cs.ucr.edu/~eamonn/time_series_data/

    work page 2015

  8. [8]

    Luke Nicholas Darlow, Qiwen Deng, Ahmed Hassan, Martin Asenov, Rajkarn Singh, Artjom Joosen, Adam Barker, and Amos Storkey. 2024. DAM: Towards a Foundation Model for Forecasting. In The Twelfth International Conference on Learning Representations

    work page 2024

Show all 49 references
  1. [9]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. [n. d.]. A decoder- only foundation model for time-series forecasting. In Forty-first International Conference on Machine Learning

  2. [10]

    Jinliang Deng, Xiusi Chen, Renhe Jiang, Du Yin, Yi Yang, Xuan Song, and Ivor W Tsang. 2024. Disentangling structured components: Towards adaptive, inter- pretable and scalable time series forecasting. IEEE Transactions on Knowledge and Data Engineering (2024)

    2024

  3. [11]

    Jinliang Deng, Feiyang Ye, Du Yin, Xuan Song, Ivor Tsang, and Hui Xiong. 2024. Parsimony or capability? decomposition delivers both in long-term time series forecasting. Advances in Neural Information Processing Systems 37 (2024), 66687– 66712

    2024

  4. [12]

    Samuel Dooley, Gurnoor Singh Khurana, Chirag Mohapatra, Siddartha Naidu, and Colin White. 2023. ForecastPFN: synthetically-trained zero-shot forecasting. In NeurIPS

    2023

  5. [13]

    Gifford, Chandra Reddy, and Jayant Kalagnanam

    Vijay Ekambaram, Arindam Jati, Pankaj Dayama, Sumanta Mukherjee, Nam H Nguyen, Wesley M. Gifford, Chandra Reddy, and Jayant Kalagnanam. 2024. Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series. In NeurIPS

    2024

  6. [14]

    Shanghua Gao, Teddy Koker, Owen Queen, Thomas Hartvigsen, Theodoros Tsiligkaridis, and Marinka Zitnik. 2024. UniTS: A Unified Multi-Task Time Series Model. In NeurIPS

    2024

  7. [15]

    Azul Garza and Max Mergenthaler-Canseco. 2023. TimeGPT-1. arXiv preprint arXiv:2310.03589 (2023)

    2023 arXiv

  8. [16]

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

    2021 arXiv

  9. [17]

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. [n. d.]. MOMENT: A Family of Open Time-series Foundation Models. In Forty-first International Conference on Machine Learning

  10. [18]

    Jincai Huang, Yongjun Xu, Qi Wang, Qi Cheems Wang, Xingxing Liang, Fei Wang, Zhao Zhang, Wei Wei, Boxuan Zhang, Libo Huang, et al. 2025. Foundation models and intelligent decision-making: Progress, challenges, and perspectives. The Innovation (2025)

    2025

  11. [19]

    Chenxi Liu, Hao Miao, Qianxiong Xu, Shaowen Zhou, Cheng Long, Yan Zhao, Ziyue Li, and Rui Zhao. 2025. Efficient Multivariate Time Series Forecasting via Calibrated Language Models with Privileged Knowledge Distillation. In ICDE

    2025

  12. [20]

    Chenxi Liu, Qianxiong Xu, Hao Miao, Sun Yang, Lingzheng Zhang, Cheng Long, Ziyue Li, and Rui Zhao. 2025. Timecma: Towards llm-empowered multivariate time series forecasting via cross-modality alignment. In AAAI, Vol. 39

    2025

  13. [21]

    Xu Liu, Juncheng Liu, Gerald Woo, Taha Aksu, Yuxuan Liang, Roger Zimmer- mann, Chenghao Liu, Silvio Savarese, Caiming Xiong, and Doyen Sahoo. 2024. Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts. arXiv preprint arXiv:2410.10469 (2024)

    2024 arXiv

  14. [22]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-1...

    2024

  15. [23]

    Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, and Mingsheng Long

  16. [24]

    Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. [n. d.]. Timer: Generative Pre-trained Transformers Are Large Time Series Models. In Forty-first International Conference on Machine Learning

  17. [25]

    Zhiding Liu, Jiqian Yang, Mingyue Cheng, Yucong Luo, and Zhi Li. 2024. Genera- tive pretrained hierarchical transformer for time series forecasting. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2024

  18. [26]

    Andrzej Maćkiewicz and Waldemar Ratajczak. 1993. Principal components analysis (PCA). Computers & Geosciences 19, 3 (1993), 303–342

    1993

  19. [27]

    Spyros Makridakis, Evangelos Spiliotis, and Vassilios Assimakopoulos. 2022. M5 accuracy competition: Results, findings, and conclusions. International Journal of Forecasting 38, 4 (2022), 1346–1364

    2022

  20. [28]

    Leland McInnes and John Healy. 2018. UMAP: Uniform Manifold Approxi- mation and Projection for Dimension Reduction. CoRR abs/1802.03426 (2018). arXiv:1802.03426

    2018 arXiv

  21. [29]

    Hao Miao, Ziqiao Liu, Yan Zhao, Chenjuan Guo, Bin Yang, Kai Zheng, and Christian S Jensen. 2024. Less is more: Efficient time series dataset condensation via two-fold modal matching. PVLDB 18, 2 (2024), 226–238

    2024

  22. [30]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In ICLR. OpenReview.net

    2023

  23. [31]

    OpenAI. 2023. GPT-4 Technical Report. CoRR abs/2303.08774 (2023). arXiv:2303.08774

    2023 arXiv

  24. [32]

    John Paparrizos, Yuhao Kang, Paul Boniol, Ruey S Tsay, Themis Palpanas, and Michael J Franklin. 2022. Tsb-uad: an end-to-end benchmark suite for univariate time-series anomaly detection. Proceedings of the VLDB Endowment 15, 8 (2022), 1697–1711

    2022

  25. [33]

    Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S Jensen, Zhenli Sheng, et al. 2024. Tfb: Towards comprehensive and fair benchmarking of time series forecasting meth- ods. arXiv preprint arXiv:2403.20150 (2024)

    2024 arXiv

  26. [34]

    Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Arian Khorasani, George Adamopoulos, Rishika Bhagwatkar, Marin Biloš, Hena Ghonia, Nadhir Hassen, Anderson Schneider, et al. 2023. Lag-llama: Towards foundation models for time series forecasting. In R0-FoMo: Robustness of Few...

    2023

  27. [35]

    Yunfan Shao, Linyang Li, Zhaoye Fei, Hang Yan, Dahua Lin, and Xipeng Qiu

  28. [36]

    Zezhi Shao, Tangwen Qian, Tao Sun, Fei Wang, and Yongjun Xu. 2025. Spatial- temporal large models: A super hub linking multiple scientific areas with artificial intelligence. The Innovation 6, 2 (2025)

    2025

  29. [37]

    arXiv preprint arXiv:2402.14526 (2024)

    Balanced Data Sampling for Language Model Training with Clustering. arXiv preprint arXiv:2402.14526 (2024)

    2024 arXiv

  30. [38]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lam- ple. 2023. LLaMA: Open and Efficient Foundation ...

    2023 arXiv

  31. [39]

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. 2024. Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts. arXiv preprint arXiv:2409.16040 (2024)

    2024 arXiv

  32. [40]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In NeurIPS

    2017

  33. [41]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008)

    2008

  34. [42]

    Yihang Wang, Yuying Qiu, Peng Chen, Kai Zhao, Yang Shu, Zhongwen Rao, Lujia Pan, Bin Yang, and Chenjuan Guo. 2024. ROSE: Register Assisted General Time Series Forecasting with Decomposed Frequency Learning. arXiv preprint arXiv:2405.17478 (2024)

    2024 arXiv

  35. [43]

    Chengsen Wang, Qi Qi, Jingyu Wang, Haifeng Sun, Zirui Zhuang, Jinming Wu, Lei Zhang, and Jianxin Liao. 2025. ChatTime: A Unified Multimodal Time Series Foundation Model Bridging Numerical and Textual Data. In AAAI Conference on Artificial Intelligence

    2025

  36. [44]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2023. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. In ICLR. OpenReview.net

    2023

  37. [45]

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. [n. d.]. Unified Training of Universal Time Series Forecasting Transformers. In Forty-first International Conference on Machine Learning

  38. [46]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are Transformers Effective for Time Series Forecasting? AAAI

    2023

  39. [47]

    Haixu Wu, Hang Zhou, Mingsheng Long, and Jianmin Wang. 2023. Interpretable weather forecasting for worldwide stations with a unified deep model. Nature Machine Intelligence 5, 6 (2023), 602–611

    2023

  40. [49]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. In AAAI. 11106–11115. BLAST: Balanced Sampling Time Series Corpus for Universal Forecasting Mo...

    2021

  41. [2024]

    arXiv preprint arXiv:2410.04803 (2024)

    Timer-XL: Long-Context Transformers for Unified Time Series Forecasting. arXiv preprint arXiv:2410.04803 (2024)

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.