Pith. sign in

REVIEW 6 minor 71 references

This review argues that vector autoregressions, transformers, foundation models, and diffusion forecasters are all attempts to answer one question—what is the conditional law of the future given the past—and that modern methods gain flexibi

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

AI forecasting methods are flexible generalizations of the classical VAR's conditional forecast distribution, gaining adaptability and scale but losing ready-made inference, identification, and structural interpretation.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A genuinely useful review that makes the attention–VAR analogy precise with Eq. (5)–(6), hedges where it should, and has only minor soft spots (a dated GIFT-Eval snapshot, a duplicated sentence).

arxiv 2607.14279 v1 pith:5UNXAOOU submitted 2026-07-15 econ.EM

From Vector Autoregressions to AI-based Time Series Forecasting: A Review

Likai Chen , Weining Wang This is my paper
classification econ.EM MSC 62M1091B84
keywords vector autoregressiontransformersfoundation modelsdenoising diffusionprobabilistic forecastinghigh-dimensional time seriesnonstationarityconditional predictive distribution
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 tries to establish that the recent turn to AI-based forecasting is not a break from classical econometrics but an expansion of its template. Every model in this review—VARs, transformers, pretrained foundation models, and diffusion generators—is read as an answer to the same question: what is the conditional distribution of the future segment given the observed history and covariates? The differences are differences in flexibility: attention weights replace fixed lag coefficients, pretraining replaces per-data-set estimation, and sampling replaces a closed-form predictive density. Because these methods represent the same object in different ways, the review's Table 4 compares them on a common ledger, concluding that modern methods often forecast well but lack the fixed coefficients, identification, and sampling theory that support impulse responses, Granger tests, and policy analysis. A sympathetic reader cares because this recasts the debate as a division of labour and identifies concrete open problems where econometrics remains necessary.

Core claim

The central claim, on the paper's own terms, is that the forecasting object is the same across method families: the conditional law of the future segment given the observed history and covariates. A Gaussian VAR represents this law with fixed linear coefficients and a Gaussian predictive density; a transformer represents it with data-dependent softmax attention weights over transformed lags; a foundation model conditions a pretrained network on the supplied history without re-estimation; a diffusion forecaster samples joint future paths from the predictive law. Equation (5) is the anchor: a one-step transformer forecast is a weighted average of transformed past values, with attention weights

What carries the argument

The carrying object is the conditional distribution of the future given the past, used as a common yardstick that lets the review place VARs and modern AI models on one spectrum. The specific mechanism that carries the argument is the formal correspondence between a VAR's fixed lag coefficients and a transformer's attention weights: equation (5) shows a forecast as a weighted sum of transformed lags with input-dependent softmax weights. This correspondence does the work of locating the gain (adaptivity and nonlinearity) and the loss (the fixed coefficients on which impulse responses, Granger causality, and sampling distributions depend). Supporting machinery includes the three challenges—hig

Load-bearing premise

The argument assumes that VARs, transformers, foundation models, and diffusion models are estimating the same object—the conditional distribution of the future given the past—rather than different kinds of objects; if that premise fails, the claim that they differ only in flexibility, and that inference is what was lost, does not follow.

What would settle it

A direct test: fit a channel-independent patch transformer and a VAR(1) to a stationary bivariate Gaussian process with strong cross-sectional correlation (say 0.9 at lag 0). The transformer's predictive law will be the product of the two marginals; if that product does not equal the true joint conditional law, then the two methods are not representing the same object, and the unifying-view premise fails for this architecture family.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Attention weights in a fitted transformer are not fixed coefficients; they should not be read as impulse responses or given structural interpretations without additional identification.
  • Zero-shot foundation-model results must be judged on contamination-aware, post-cutoff data; leaderboard scores are dated snapshots, not stable model properties.
  • Diffusion forecasters can sample scenario-coherent future paths, but they do not enforce long-run equilibrium; cointegrated systems still need a vector error-correction guardrail.
  • Normalization-based handling of nonstationarity is a device rather than a model; unit roots and cointegration remain the econometric tools required when long-run structure matters.
  • The review's bottom line for practice: evaluate modern forecasters against strong classical baselines—seasonal naive, ARIMA, Bayesian VAR, VECM—and check interval calibration rather than assume it.

Where Pith is reading between the lines

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

  • The unifying-view thesis implies a concrete research program: characterize when pretrained in-context learning transfers from i.i.d. task mixtures to a single nonstationary dependent path—an extension the review flags as open but does not develop.
  • If attention weights are learnable kernels, they could be probed like fitted smoothers—tracking effective lag profiles across regimes—offering a diagnostic test of the adaptivity claim even before full inference is available.
  • The asymmetry thesis suggests hybrid architectures are the natural frontier: generative heads or transformer representations attached to structural backbones, so that sampled scenarios respect identified shocks and equilibrium-restoring terms; the review gestures at this convergence but leaves the design open.
  • A quantitative prediction follows from the paper's own premise: on cointegrated DGPs, normalization-based forecasters should show growing long-horizon error relative to a VECM as the error-correction strength increases; this is directly testable on synthetic data.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 6 minor

Summary. This review surveys the connection between classical VAR-based econometric forecasting and three recent AI-based families: transformers, pretrained foundation models for zero-shot forecasting, and diffusion-based generative forecasters. It proposes a unifying object—the conditional law of the future given the past—and argues that the modern methods are best read as more flexible representations of that same object, while giving up the fixed-coefficient structure that supports impulse-response analysis, Granger-causality tests, cointegration, and classical sampling inference. The discussion is organized around three challenges: high dimensionality, nonstationarity, and nonlinearity. The closing synthesis argues that the remaining agenda—inference, identification, and nonstationarity—is precisely where econometric tools retain their value.

Significance. If accepted, the paper provides a genuinely useful bridge between the econometric and machine-learning forecasting literatures. It is a review, not an original theoretical contribution, but it is a fair and accurate synthesis. Its strengths are a clear organizing framework, an honest treatment of benchmark evidence and contamination, a standard and correctly presented classical econometrics section, and explicit statements of the limitations of modern methods. No machine-checked proofs or reproducible code are supplied, but neither is expected for a survey of this kind. The central asymmetry claim—modern methods forecast well but often lack structural and inferential tools—is a defensible survey-level interpretation rather than a precise theorem, and the paper's own hedging is appropriate.

minor comments (6)
  1. [§3.2, Eq. (5)] The display equation is dimensionally inconsistent as written: y_t is in R^p, but the right-hand side is a weighted average of value vectors in R^{d_v}, and no output projection, residual connection, or normalization is shown. The text says Eq. (5) "states the correspondence precisely"; it would be more accurate to call it a schematic representation and to include the output projection or explicitly restrict attention to a stylized single-layer case.
  2. [Table 2] The table is well documented as a dated snapshot, but for reproducibility it would help to provide the exact GIFT-Eval repository version or commit hash and the retrieval date in the caption or footnote. The three-decimal precision suggests a stability that a changing leaderboard may not support; rounding to two decimals would be more honest.
  3. [Table 1, Time-MoE row] The tokenization category for Time-MoE should be verified. The cited model is commonly described as using multi-resolution patching rather than point-wise real values. If so, the table entry should be updated to avoid a factual misclassification.
  4. [Figure 4 caption] The phrase "real-valued patches are compact and probabilistic by construction" is unclear. Real-valued patches are continuous-valued representations; they are not probabilistic unless a distributional output head is added. Please rephrase.
  5. [§5, before 5.1] The sentence "For nonstationarity, the picture is less settled." appears twice consecutively. Delete the duplicate.
  6. [§3.2] The analogy between attention weights and VAR coefficients would be more precise if the paper noted that attention weights are nonnegative and sum to one, whereas VAR coefficients are unconstrained. The learned value transform partially offsets this difference, but the restriction is worth stating explicitly.

Circularity Check

0 steps flagged

No significant circularity: this is an interpretive survey whose claims rest on external literature and public benchmarks, with no prediction derived from its own inputs.

full rationale

The paper makes no original derivation whose conclusion is built into its inputs. Its central organizing claim—that all reviewed models can be read as answering the question of the conditional law of the future given the past—is explicitly framed as an interpretive device rather than a theorem: 'Every model in this review can be read as an answer to a single question' (Section 1). The formal correspondence between attention and VAR in Eq. (5) is stated with its own limitations acknowledged: 'Equation (5) is a weighted average of transformed lags... the analogy has two limits'—fixed versus data-dependent weights and linearity versus nonlinearity. Similarly, the Wold/diffusion parallel is explicitly marked as merely formal: 'The resemblance is formal rather than substantive.' The empirical claim about foundation models is attributed to an external benchmark (GIFT-Eval), and the paper itself flags the reproduction caveat: 'The leaderboard is updated over time, so these values should be read as a dated snapshot' and warns that contamination 'is hard to verify.' The asymmetry claim—modern methods forecast well but lack inferential/structural tools—is a descriptive comparison of model features (absence of impulse responses, testable restrictions, sampling distributions), not a derived equivalence. No parameter is fitted to data and then relabeled as a prediction, no cited result from the authors' own prior work is used to force a conclusion, and no uniqueness theorem is imported from self-citation. The paper is self-contained as a review, with all substantive content attributed to cited external work. Accordingly, no circular step is present, and the honest finding is a score of 0.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No free parameters or invented entities: this is a review. The load-bearing axioms are the unifying conditional-law premise and the GIFT-Eval benchmark premise; both are the authors' organizational and empirical assumptions, not facts proven here.

axioms (4)
  • domain assumption All methods can be represented as estimators of the same conditional law p(y_{t+1:t+H} | y_{1:t}).
    Section 1 'A unifying view'; the entire comparison of gains and losses depends on this identification. If false, the flexibility-vs-inference ledger is not well defined.
  • domain assumption The GIFT-Eval snapshot accurately reflects current zero-shot forecasting performance.
    Table 2 and the claim that modern methods 'often forecast well' rest on this; the paper itself flags contamination and snapshot instability, so this premise is load-bearing and uncertain.
  • domain assumption Impulse responses and Granger tests require fixed coefficients; input-dependent attention weights have no classical sampling distribution.
    Sections 2 and 3.2; this is the basis for the claimed loss of inferential and structural tools. It is conventional but a stated modeling premise.
  • ad hoc to paper Equation (5) is a legitimate formal correspondence between attention and a VAR with data-dependent weights.
    Section 3.2; the paper constructs this representation by setting x_s = E(y_s) and taking the query from the last token. It is an expository device, not a derived theorem.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of From Vector Autoregressions to AI-based Time Series Forecasting: A Review." pith.science (2026). https://pith.science/paper/5UNXAOOU

@misc{pith2026260714279,
  author       = {Pith},
  title        = {Pith review of: From Vector Autoregressions to AI-based Time Series Forecasting: A Review},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5UNXAOOU}},
  note         = {Machine review of arXiv:2607.14279}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Forecasting is a central goal of time-series analysis. This review centers on three major developments in recent AI-based time-series forecasting: transformers, large pretrained models for zero-shot forecasting, and diffusion-based generative forecasters. We connect these methods to the econometric tradition built around the vector autoregression (VAR) through a common object: the conditional distribution of the future given the past. The review is organized around three long-standing challenges: \emph{high dimensionality}, \emph{nonstationarity}, and \emph{nonlinearity}. We argue that modern methods make progress by expanding the classical forecasting template: they allow more flexible dynamics, use larger information sets and training corpora, and represent richer predictive distributions. Yet they often lack the inferential and structural tools that make classical models useful for testing, explanation, and policy analysis. We close by outlining open problems where econometric tools remain important.

Figures

Figures reproduced from arXiv: 2607.14279 by Likai Chen, Weining Wang.

Figure 1
Figure 1. Figure 1: The central relaxation of this review. Left: the VAR forecasts with fixed coefficient [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Self-attention as a learned weighted estimator. The input [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Tokenization in language modelling. A sentence is segmented into subword tokens, [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Two tokenization schemes for pretrained forecasters. Left: each scaled value is quan [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Why generative forecasting. A point forecast is a single trajectory and conveys little [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The diffusion mechanism. A fixed forward process gradually adds noise to the series [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Autoregressive versus whole-segment generation. TimeGrad generates one date at a [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Multiscale structure in diffusion. MG-TSD uses coarse-grained series as guidance [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

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

71 extracted references · 6 linked inside Pith

  1. [1]

    GIFT-Eval: A benchmark for general time series forecasting model evaluation.arXiv preprint arXiv:2410.10393, 2024

    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

    Pith/arXiv arXiv 2024

  2. [2]

    Diffusion-based time series imputation and forecasting with structured state space models.Transactions on Machine Learning Research (TMLR), 2023

    Juan Miguel Lopez Alcaraz and Nils Strodthoff. Diffusion-based time series imputation and forecasting with structured state space models.Transactions on Machine Learning Research (TMLR), 2023

    2023

  3. [3]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Ali Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Ma- honey, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Bernie Wang. Chronos: Learning the lan...

  4. [4]

    Chronos-2: From univariate to universal forecasting.arXiv preprint arXiv:2510.15821, 2025

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

    Pith/arXiv arXiv 2025

  5. [5]

    Determining the number of factors in approximate factor models

    Jushan Bai and Serena Ng. Determining the number of factors in approximate factor models. Econometrica, 70(1):191–221, 2002

    2002

  6. [6]

    Large Bayesian vector auto regressions.Journal of Applied Econometrics, 25(1):71–92, 2010

    Marta Bańbura, Domenico Giannone, and Lucrezia Reichlin. Large Bayesian vector auto regressions.Journal of Applied Econometrics, 25(1):71–92, 2010

    2010

  7. [7]

    Regularized estimation in sparse high-dimensional time series models.The Annals of Statistics, 43(4):1535–1567, 2015

    Sumanta Basu and George Michailidis. Regularized estimation in sparse high-dimensional time series models.The Annals of Statistics, 43(4):1535–1567, 2015

    2015

  8. [8]

    The dynamic effects of aggregate demand and supply disturbances.The American Economic Review, 79(4):655–673, 1989

    Olivier Jean Blanchard and Danny Quah. The dynamic effects of aggregate demand and supply disturbances.The American Economic Review, 79(4):655–673, 1989

    1989

  9. [9]

    Olivares, Boris N

    Cristian Challu, Kin G. Olivares, Boris N. Oreshkin, Federico Garza Ramirez, Max Mergen- thaler Canseco, and Artur Dubrawski. NHITS: Neural hierarchical interpolation for time series forecasting.Proceedings of the AAAI Conference on Artificial Intelligence, 37(6): 6989–6997, 2023

    2023

  10. [10]

    Fitting time series models to nonstationary processes.The Annals of Statistics, 25(1):1–37, 1997

    Rainer Dahlhaus. Fitting time series models to nonstationary processes.The Annals of Statistics, 25(1):1–37, 1997

    1997

  11. [11]

    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. InInternational Conference on Machine Learning (ICML), 2024

    2024

  12. [12]

    Dickey and Wayne A

    David A. Dickey and Wayne A. Fuller. Distribution of the estimators for autoregressive time series with a unit root.Journal of the American Statistical Association, 74(366):427–431, 1979

    1979

  13. [13]

    Engle and C

    Robert F. Engle and C. W. J. Granger. Co-integration and error correction: Representation, estimation, and testing.Econometrica, 55(2):251–276, 1987

    1987

  14. [14]

    MG-TSD: Multi-granularity time series diffusion models with guided learning process

    Xinyao Fan, Yueying Wu, Chang Xu, Yuhao Huang, Weiqing Liu, and Jiang Bian. MG-TSD: Multi-granularity time series diffusion models with guided learning process. InInternational Conference on Learning Representations (ICLR), 2024. 22

    2024

  15. [15]

    Wild binary segmentation for multiple change-point detection.The Annals of Statistics, 42(6):2243–2281, 2014

    Piotr Fryzlewicz. Wild binary segmentation for multiple change-point detection.The Annals of Statistics, 42(6):2243–2281, 2014

    2014

  16. [16]

    TimeGPT-1.arXiv preprint arXiv:2310.03589, 2023

    Azul Garza, Cristian Challu, and Max Mergenthaler-Canseco. TimeGPT-1.arXiv preprint arXiv:2310.03589, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    C. W. J. Granger. Investigating causal relations by econometric models and cross-spectral methods.Econometrica, 37(3):424–438, 1969

    1969

  18. [18]

    Granger and P

    C.W.J. Granger and P. Newbold. Spurious regressions in econometrics.Journal of Econo- metrics, 2(2):111–120, 1974. ISSN 0304-4076

    1974

  19. [19]

    Efficiently modeling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  20. [20]

    Hamilton.Time Series Analysis

    James D. Hamilton.Time Series Analysis. Princeton University Press, 1994

    1994

  21. [21]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  22. [22]

    Hyndman and Yeasmin Khandakar

    Rob J. Hyndman and Yeasmin Khandakar. Automatic time series forecasting: The forecast package for R.Journal of Statistical Software, 27(3):1–22, 2008

    2008

  23. [23]

    Estimation and hypothesis testing of cointegration vectors in gaussian vector autoregressive models.Econometrica, 59(6):1551–1580, 1991

    Søren Johansen. Estimation and hypothesis testing of cointegration vectors in gaussian vector autoregressive models.Econometrica, 59(6):1551–1580, 1991

    1991

  24. [24]

    Oxford University Press, 1995

    Søren Johansen.Likelihood-Based Inference in Cointegrated Vector Autoregressive Models. Oxford University Press, 1995

    1995

  25. [25]

    Cambridge University Press, 2017

    Lutz Kilian and Helmut Lütkepohl.Structural Vector Autoregressive Analysis. Cambridge University Press, 2017

    2017

  26. [26]

    Reversible instance normalization for accurate time-series forecasting against distribution shift

    Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. Reversible instance normalization for accurate time-series forecasting against distribution shift. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  27. [27]

    Oracle inequalities for high dimensional vector autoregressions.Journal of Econometrics, 186(2):325–344, 2015

    Anders Bredahl Kock and Laurent Callot. Oracle inequalities for high dimensional vector autoregressions.Journal of Econometrics, 186(2):325–344, 2015

    2015

  28. [28]

    Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting

    Marcel Kollovieh, Abdul Fatir Ansari, Michael Bohlke-Schneider, Jasper Zschiegner, Hao Wang, and Yuyang Wang. Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  29. [29]

    Ant: Adaptive noise schedule for time series diffusion models

    Seunghan Lee, Kibok Lee, and Taeyoung Park. Ant: Adaptive noise schedule for time series diffusion models. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  30. [30]

    Automatic change-point detection in time series via deep learning.Journal of the Royal Statistical Society Series B: Statistical Methodology, 86(2):273–285, 2024

    Jie Li, Paul Fearnhead, Piotr Fryzlewicz, and Tengyao Wang. Automatic change-point detection in time series via deep learning.Journal of the Royal Statistical Society Series B: Statistical Methodology, 86(2):273–285, 2024

    2024

  31. [31]

    Generative time series forecasting with diffusion, denoise, and disentanglement

    Yan Li, Xinjiang Lu, Yaqing Wang, and Dejing Dou. Generative time series forecasting with diffusion, denoise, and disentanglement. InAdvances in Neural Information Processing Systems (NeurIPS), 2022. 23

    2022

  32. [32]

    Transformer-modulated diffusion models for probabilistic multivariate time series forecast- ing

    Yuxin Li, Wenchao Chen, Xinyue Hu, Bo Chen, Baolin Sun, and Mingyuan Zhou. Transformer-modulated diffusion models for probabilistic multivariate time series forecast- ing. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  33. [33]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  34. [34]

    Litterman

    Robert B. Litterman. Forecasting with Bayesian vector autoregressions—five years of expe- rience.Journal of Business & Economic Statistics, 4(1):25–38, 1986

    1986

  35. [35]

    Moirai 2.0: When less is more for time series forecasting.arXiv preprint arXiv:2511.11698, 2025

    Chenghao Liu, Taha Aksu, Juncheng Liu, Xu Liu, Hanshu Yan, Quang Pham, Silvio Savarese, Doyen Sahoo, Caiming Xiong, and Junnan Li. Moirai 2.0: When less is more for time series forecasting.arXiv preprint arXiv:2511.11698, 2025

    arXiv 2025

  36. [36]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InInternational Conference on Learning Representa- tions (ICLR), 2023

    2023

  37. [37]

    Non-stationary transformers: Exploring the stationarity in time series forecasting

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Non-stationary transformers: Exploring the stationarity in time series forecasting. InAdvances in Neural Information Processing Systems, 2022

    2022

  38. [38]

    iTransformer: Inverted transformers are effective for time series forecasting

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. iTransformer: Inverted transformers are effective for time series forecasting. In International Conference on Learning Representations (ICLR), 2024

    2024

  39. [39]

    Springer, 2005

    Helmut Lütkepohl.New Introduction to Multiple Time Series Analysis. Springer, 2005

    2005

  40. [40]

    Stock, and Mark W

    Massimiliano Marcellino, James H. Stock, and Mark W. Watson. A comparison of direct and iterated multistep AR methods for forecasting macroeconomic time series.Journal of Econometrics, 135(1–2):499–526, 2006

    2006

  41. [41]

    Caspar Meijer and Lydia Y. Chen. The rise of diffusion models in time-series forecasting. arXiv preprint arXiv:2401.03006, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Transformers can do Bayesian inference

    Samuel Müller, Noah Hollmann, Sebastian Pineda Arango, Josif Grabocka, and Frank Hut- ter. Transformers can do Bayesian inference. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  43. [43]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  44. [44]

    Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio

    Boris N. Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio. N-BEATS: Neural basis expansion analysis for interpretable time series forecasting. InInternational Conference on Learning Representations (ICLR), 2020

    2020

  45. [45]

    Peter C. B. Phillips and Pierre Perron. Testing for a unit root in time series regression. Biometrika, 75(2):335–346, 1988

    1988

  46. [46]

    Autoregressive denois- ing diffusion models for multivariate probabilistic time series forecasting

    Kashif Rasul, Calvin Seward, Ingmar Schuster, and Roland Vollgraf. Autoregressive denois- ing diffusion models for multivariate probabilistic time series forecasting. InInternational Conference on Machine Learning (ICML), 2021. 24

    2021

  47. [47]

    Lag-llama: Towards foundation models for probabilistic time series forecasting.arXiv preprint arXiv:2310.08278, 2023

    Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Hena Ghonia, Rishika Bhagwatkar, Arian Khorasani, Mohammad Javad Darvishi Bayazi, George Adamopoulos, Roland Riachi, Nadhir Hassen, Marin Biloš, Sahil Garg, Anderson Schneider, Nicolas Chapados, Alexandre Drouin, Valentina Zantedeschi, Yuriy Nevmyvaka, and Irina Rish. Lag-llama: Towards foundation models ...

    Pith/arXiv arXiv 2023

  48. [48]

    DeepAR: Prob- abilistic forecasting with autoregressive recurrent networks.International Journal of Fore- casting, 36(3):1181–1191, 2020

    David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. DeepAR: Prob- abilistic forecasting with autoregressive recurrent networks.International Journal of Fore- casting, 36(3):1181–1191, 2020

    2020

  49. [49]

    Lifeng Shen and James T. Kwok. Non-autoregressive conditional diffusion models for time series prediction. InInternational Conference on Machine Learning (ICML), 2023

    2023

  50. [50]

    Lifeng Shen, Weiyu Chen, and James T. Kwok. Multi-resolution diffusion models for time series forecasting. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  51. [51]

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

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

    2025

  52. [52]

    Christopher A. Sims. Macroeconomics and reality.Econometrica, 48(1):1–48, 1980

    1980

  53. [53]

    Weiss, Niru Maheswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InProceedings of the 32nd International Conference on Machine Learning (ICML), volume 37 ofPMLR, pages 2256– 2265, 2015

    2015

  54. [54]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021

    2021

  55. [55]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. InAdvances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  56. [56]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  57. [57]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Proceedings of the 40th International Conference on Machine Learning (ICML), 2023

    2023

  58. [58]

    Stock and Mark W

    James H. Stock and Mark W. Watson. Forecasting using principal components from a large number of predictors.Journal of the American Statistical Association, 97(460):1167–1179, 2002

    2002

  59. [59]

    CSDI: Conditional score- based diffusion models for probabilistic time series imputation

    Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. CSDI: Conditional score- based diffusion models for probabilistic time series imputation. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  60. [60]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  61. [61]

    Transformersintimeseries: Asurvey

    Qingsong Wen, Tian Zhou, Chaoli Zhang, Weiqi Chen, Ziqing Ma, Junchi Yan, and Liang Sun. Transformersintimeseries: Asurvey. InProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI), pages 6778–6786, 2023. 25

    2023

  62. [62]

    The learnability of in-context learning

    Noam Wies, Yoav Levine, and Amnon Shashua. The learnability of in-context learning. In Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  63. [63]

    Unified training of universal time series forecasting transformers

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers. InInternational Conference on Machine Learning (ICML), 2024

    2024

  64. [64]

    Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  65. [65]

    An explanation of in-context learning as implicit Bayesian inference

    Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation of in-context learning as implicit Bayesian inference. InInternational Conference on Learning Representations (ICLR), 2022

    2022

  66. [66]

    ScoreGrad: Multi- variate probabilistic time series forecasting with continuous energy-based generative models

    Tijin Yan, Hongwei Zhang, Tong Zhou, Yufeng Zhan, and Yuanqing Xia. ScoreGrad: Multi- variate probabilistic time series forecasting with continuous energy-based generative models. arXiv preprint arXiv:2106.10121, 2021

    Pith/arXiv arXiv 2021

  67. [67]

    A survey on diffusion models for time series and spatio-temporal data.arXiv preprint arXiv:2404.18886, 2024

    Yiyuan Yang, Ming Jin, Haomin Wen, Chaoli Zhang, Yuxuan Liang, Lintao Ma, Yi Wang, Chenghao Liu, Bin Yang, Zenglin Xu, Jiang Bian, Shirui Pan, and Qingsong Wen. A survey on diffusion models for time series and spatio-temporal data.arXiv preprint arXiv:2404.18886, 2024

    arXiv 2024

  68. [68]

    Diffusion-TS: Interpretable diffusion for general time series generation

    Xinyu Yuan and Yan Qiao. Diffusion-TS: Interpretable diffusion for general time series generation. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  69. [69]

    Are transformers effective for time series forecasting? InProceedings of the AAAI Conference on Artificial Intelligence, 2023

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? InProceedings of the AAAI Conference on Artificial Intelligence, 2023

    2023

  70. [70]

    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. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021

    2021

  71. [71]

    FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. InInterna- tional Conference on Machine Learning (ICML), 2022. 26

    2022

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.