Pith. sign in

REVIEW 3 major objections 3 minor 43 references

The Spectrum Is Not Enough: When Context Helps Time-Series Forecasting

T0 review · 3 major / 3 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read How predictable a series is says nothing about whether more context helps forecast it.

desk verdict The impossibility result — no power-spectrum or autocovariance functional can decide beyond-second-order context value — is correct and worth taking seriously; the applied diagnostic, however, is not cleanly validated as defined, and the retrieval sign-prediction is partly circular. read the letter →

arxiv 2607.13006 v2 pith:B5Y4BON4 submitted 2026-07-14 cs.LG

Mert Onur Cakiroglu , Mehmet Dalkilic , Hasan Kurban This is my paper
classification cs.LG
keywords time-seriesforecastingpredictabilitypowerspectrumphaserandomizationsurrogatedataretrieval-augmentedfoundationmodelscoveragedeficit
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 tries to establish that spectral predictability indices — scores computed from the power spectrum of a time series — cannot answer the deployment question of whether adding context (a longer lookback, a retrieval plug-in, or a pretrained foundation model) will improve forecasts. The reason is structural: the value of context is a property of the series plus the operating configuration (window length and horizon), not of the series alone, while every spectral index is invariant under phase randomization and the beyond-second-order structure that retrieval and foundation models exploit is not. It proves this as an impossibility theorem, isolates it with surrogate pairs that fix the spectrum and marginal by construction, and introduces a label-free pre-deployment diagnostic, the coverage deficit, whose principal term measures the analog-over-linear prediction gain. Experiments on seven benchmarks show retrieval value collapsing through zero across surrogate pairs while spectral indices freeze, matching the theory.

What carries the argument

The argument rests on three linked pieces. Theorem 1, the value ceiling, bounds any mechanism's relative error reduction over the best linear predictor by the normalized beyond-second-order gap Δ/σ²_lin; since the linear floor is an autocovariance functional, the power spectrum cannot see Δ. Second, phase-randomized surrogate pairs: randomizing Fourier phases preserves the periodogram and autocovariance (and, with amplitude adjustment, the marginal) but drives higher-order joint cumulants to zero, so the surrogate's laws converge to a Gaussian process on which Δ = 0. Third, the coverage deficit Γ = Γ_cov + Γ_oov, computed label-free before deployment: Γ_cov is Δ_nl · u(S), where Δ_nl is the

What would settle it

Fit a learned retrieval model that reads a longer context than the operating window (e.g., a trained neural retriever) on a series with measurable nonlinear structure; if its value survives phase randomization, the value ceiling no longer bounds it by the beyond-second-order gap and the surrogate-collapse dissociation in Table 1 would not hold.

Watch

Extended reading notes

Core claim

The central claim (Corollary 5) is that no predictability index that is a functional of the power spectrum or autocovariance can determine beyond-spectrum context value. Any such index assigns the same score to a series and its phase-randomized surrogate, which shares the spectrum and, after amplitude adjustment, the marginal; yet the surrogate is asymptotically Gaussian, so the gap Δ between the best linear prediction error and the Bayes error — the structure that analog retrieval and foundation-model margins exploit — vanishes on it. A series and its surrogate therefore have identical spectral scores but opposite context value whenever Δ > 0. The paper supports this with controlled experim

Load-bearing premise

The whole value of window-keyed retrieval is attributed to the beyond-second-order gap on the assumption that retrieval adds no linear predictability beyond the window itself; if a retriever — especially a learned one — enlarges the linear information set, Theorem 1 no longer bounds its value by Δ, and the surrogate collapse would not isolate the spectral blind spot.

Editorial extensions

If this is right

  • A high spectral predictability score does not license a heavier model or longer context; the decision must be made per configuration, not per series.
  • For retrieval mechanisms keyed on the operating window, the entire relative value is beyond-second-order, so it should collapse on phase-randomized surrogates; the diagnostic's structure term predicts the sign leave-one-dataset-out where spectral indices are at or below chance.
  • A longer linear lookback's value is second-order and spectrum-visible; spectral indices can rank it, and it survives phase randomization.
  • Foundation-model gains decompose: the dominant second-order part is spectrum-visible, while the smaller beyond-linear margin is invisible to spectral indices and collapses on surrogates.
  • The same impossibility applies to any newly proposed spectral index, so a phase-sensitive, configuration-level statistic is required to answer the deployment question.

Reading between the lines

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

  • If a learned retrieval model enlarges the linear information set — by reading more than the window, for example — its value would include a spectrum-visible second-order component, so the surrogate collapse would not isolate the spectral blind spot; the paper's own limitations concede this possibility.
  • The coverage-deficit rule could be tested as an online controller: a deployment system could monitor Γ_cov and Γ_oov on streaming data and switch context augmentation on or off, a use the paper does not develop.
  • For series with finitely supported spectra, Lemma 3's Lindeberg condition fails; the impossibility does not apply (such series are perfectly linearly predictable at windows longer than twice the number of spectral lines), so the diagnostic should be applied only where the condition plausibly holds.
  • The framework suggests a general recipe for any predictability index: to be useful for context decisions, an index must be a function of the configuration (window, horizon, mechanism), not just the series; this is testable by checking invariance under phase randomization at fixed configuration.
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

3 major / 3 minor

Summary. The paper distinguishes two quantities: series-level predictability indices (e.g., spectral predictability Ω, SCP) and the value of adding a context-extending mechanism M to a forecaster at a specific operating point (S,H). The theoretical core (Theorem 1, Proposition 2, Lemma 3, Proposition 4, Corollary 5) argues that any index that is a functional of the power spectrum or autocovariance is invariant under phase randomization, whereas beyond-second-order context value — the part retrieval and foundation-model margins exploit — is not. The paper then proposes a label-free, configuration-level diagnostic, the coverage deficit Γ, whose principal term is Δnl·u(S) plus a novelty term Γoov, and reports a leave-one-dataset-out sign-prediction result (Table 4) where the structure term Δnl predicts retrieval and foundation-margin sign better than spectral indices. The paper also reports a spectrum-controlled collapse: under IAAFT surrogates, retrieval value collapses (ECL +33%→−35%), spectral indices stay frozen, the longer-linear-window gain survives, and the foundation model's beyond-linear margin collapses.

Significance. If the theoretical result and the diagnostic were fully supported, the contribution would be significant: it would reframe a growing literature that uses spectral predictability scores for deployment decisions, and it would provide a concrete pre-deployment diagnostic. The theoretical part — Corollary 5 — is, under the stated assumptions, a clean and useful impossibility statement, and the surrogate-controlled E1 measurements are a well-designed positive control that visibly separates spectral invariance from context-value non-invariance. The release of code and the explicit handling of surrogate residuals are strengths. However, the paper's central applied claim — that the coverage deficit Γ(S,H) is a validated, configuration-level diagnostic — is not supported by the experiments as reported. The headline LODO accuracy is obtained from a different statistic than the one defined as the coverage deficit, and the retrieval-sign prediction is close to circular because the retrieval mechanism and the diagnostic estimate the same analog-over-linear gain. These issues are fixable within the manuscript's scope, but they require substantial reframing or additional experiments.

major comments (3)
  1. [§4, Eq. (7); Table 4; App. A.9] The headline diagnostic is not what is evaluated. Eq. (7) defines Γcov(S)=Δnl·u(S) with Δnl estimated via the default/motif embedding d=clip(L,4,32) (App. A.9), but Table 4's 'Δnl (ours)' row is the operating-window structure term with d=S=12, not Γcov. Table A1 reports Γcov=0.000 on ECL in the original arm, the same cell where retrieval value is +33% and where Table 4's Δnl reaches 78.1% LODO accuracy. App. A.10 further reports Γcov alone has 0.73 sign agreement versus 0.90 for the operating-window Δnl. Thus the paper's central applied claim — that the coverage deficit Γ(S,H) is a configuration-level diagnostic of context value — is unvalidated. Please either report LODO accuracy for the actual Γcov and for the full (Γcov,Γoov) rule, or explicitly restrict the diagnostic claim to the operating-window Δnl statistic and justify why that statistic is a configuration-level diagnostic.
  2. [§3, Eq. (5); App. A.8; §6 Table 4] The retrieval-sign result is close to circular. For window-keyed retrieval, App. A.8 states σ²_lin(I_M)=σ²_lin(I_base), so Theorem 1 gives V_retrieval ≤ Δ/σ²_lin. Eq. (5)'s Δnl is a conservative estimate of exactly Δ/σ²_lin, and the retrieval mechanism used in E1/E2 is the analog/simplex predictor whose test error is the MSE_analog of Eq. (5). Therefore Table 4's 78.1% balanced sign accuracy for Δnl on retrieval predicts the sign of an outcome from an estimate of the same outcome. This is not evidence that a diagnostic generalizes to independent context mechanisms; it is a consistency check on the estimator. The foundation-margin column weakens the circularity, but the retrieval result is the paper's flagship empirical result. To support the deployment-diagnostic claim, please evaluate with a learned retrieval mechanism (as the Limitations mention) or with a mechanism whose information s
  3. [Corollary 5; Proposition 4] Corollary 5 overstates the condition 'whenever Δ>0.' For a fixed mechanism M, V_M can be zero or negative even when Δ(I_M)>0 (for example, a retrieval key that is poorly chosen, or a mechanism that does not approach the Bayes error). The impossibility statement only requires the existence of a pair (x,x_G) with equal P and different V_M; the nonlinear-AR example in Proposition 4 supplies this. The corollary's 'whenever Δ>0' phrasing is therefore literally false for arbitrary x with Δ>0 and should be replaced by an existence/attainability statement. The central conclusion can be repaired, but the current wording needs correction.
minor comments (3)
  1. [§5, E1; Table A1] The protocol excludes cells failing the periodogram-residual gate, but the number and identity of excluded cells are not reported. Please report the exclusion count and verify that the E1/E2 conclusions are unchanged if the gate is tightened or loosened.
  2. [§6, Table 3] The foundation-model 'margin' is defined over a train-fit linear predictor, not the population best linear predictor on the same access. The text calls it the empirical counterpart of Theorem 1, but this should be stated more prominently to avoid confusion between a finite-sample quantity and the theorem's population bound.
  3. [§8, Limitations] The Limitations honestly note that a learned SOTA retriever may add linear predictability and that deep SOTA backbones remain untested. These are substantive scope restrictions on the deployment-diagnostic claim; they should be reflected in the Introduction and Conclusion, not only in the Limitations section.

Circularity Check

1 steps flagged · score 6.0 of 10

The retrieval sign prediction is largely self-confirming: Δnl is the same analog-over-linear gain as V_M for window-keyed retrieval, though the FM-margin and longer-lookback columns add independent content.

  1. self definitional [Eq. (1) and Eq. (5); Sec. 4; App. A.9; Table 4]
    "V_M(x;S,H)= MSE(f)−MSE(f⊕M)/MSE(f) ... Δnl =1− MSEanalog/MSElinear ... The primary E2/E4 rule (Table 4) is the operating-window structure term: Δnl with delay-embedding dimension d=S ... analog retrieval keyed on the S-window over the training record (the simplex predictor of Sugihara and May, 1990; adds no linear information)."

    For window-keyed retrieval, f⊕M is the simplex/nearest-neighbor predictor and f is the least-squares predictor, so V_M is the normalized analog-over-linear gain at (S,H). Eq. (5) defines Δnl as the same normalized analog-over-linear gain (one-step, with d=S in the primary rule). Using Δnl to predict the sign of retrieval value is therefore predicting the sign of the quantity that Δnl itself estimates; the LODO threshold fit does not add independence. The foundation-margin and longer-lookback columns are not equivalent to Δnl, so the circularity is partial.

full rationale

The impossibility theorem (Corollary 5) is not circular: it follows from spectral invariance (Prop. 2) and the Gaussian-endpoint calculation (Prop. 4) that a power-spectrum functional is constant on (x, x_G) while the normalized gap V̄ differs; no fitted outcome is involved. The circularity is in the applied retrieval claim. Eq. (1) measures retrieval value as the gain of an analog (simplex) predictor over a linear predictor on the S-window; Eq. (5)'s Δnl is the same analog-over-linear gain (one-step, operating-window d=S in Table 4). Thus the 78.1% LODO retrieval accuracy largely predicts the sign of the very quantity the statistic estimates, a self-definitional validation rather than an external check. The FM-margin and longer-window columns are genuinely separate, which prevents a higher score. Per the reviewing rule, I also flag a validation mismatch, though it is a support/correctness issue rather than a circularity: Table A1 reports Γcov=0.000 on the key ECL original arm, while Table 4's success uses the operating-window Δnl (App. A.9), so the coverage deficit as defined in Eq. (7) is not directly validated by the headline experiment.

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

The central claim rests on standard surrogate-data theory, the linear floor being an autocovariance functional, and an attribution assumption that window-keyed retrieval adds no linear information. The E2 retrieval prediction is further tied to the analog-vs-linear gain, making the diagnostic partly self-referential for that mechanism. No new physical entities are postulated; the coverage deficit is a statistic, not an entity.

free parameters (5)
  • nearest-neighbor count k in Δnl = 4
    Chosen by hand (App. A.9); the ablation shows V_M sign is stable across k=2,4,8, but the level of Δnl depends on k.
  • embedding dimension d for Δnl = default d=clip(L,4,32); E2 rule uses d=S=12
    Two variants are used: the 'motif' variant in App. A.9 and the 'operating-window' variant that actually produces Table 4. The choice changes whether Δnl is zero on ECL or signal-carrying.
  • motif length m (default dominant period L) = per-channel periodogram peak L
    Used in the coverage fraction u(S) in Eq. (6); it is a proxy for the state-identifying motif length and is acknowledged in Limitations to degrade on multi-scale states.
  • decision-rule threshold and direction on (Γcov, Γoov) = fit on six training datasets by balanced accuracy
    The sign predictor is thresholded/logistic with parameters tuned on the development datasets before LODO evaluation (Sec. 5, E2). This is a fitted decision constant.
  • IAAFT gate, surrogate count K, novelty bins b = gate 0.02; K=20 (10 for FM); b unspecified
    Quality-control and protocol choices: cells failing the periodogram residual gate are excluded; K affects the median gap slightly; the Γoov discretization bin count b is not stated.
assumptions (5)
  • domain assumption The series x is real, second-order stationary, and finite-variance.
    Stated in Sec. 3 and repeated in Limitations; the autocovariance/spectrum representation and the Gaussian endpoint depend on it. Nonstationarity is only partially absorbed by z-normalization and the ADF gate.
  • domain assumption Lindeberg no-dominant-line condition for phase-randomized Gaussianization (max_k a_k^2 / s_T^2 → 0).
    Lemma 3 and App. A.5 require this to kill higher cumulants; finitely supported spectra are explicitly excluded. ECL and Traffic have strong daily/weekly periodic components that may approach the excluded regime.
  • domain assumption Condition (M): conditional means converge in L2 for the full Bayes-gap closure.
    App. A.5 states it explicitly and says no main-text claim uses it. It is needed only if one wants the full Δ→0, not just fixed-degree polynomial predictors.
  • domain assumption Window-keyed retrieval adds no linear predictability: σ²_lin(I_M) = σ²_lin(I_base).
    App. A.8; this is load-bearing for attributing retrieval's entire value to the beyond-second-order gap and for interpreting the surrogate collapse as erasing that value.
  • domain assumption IAAFT surrogates preserve the periodogram up to a residual and the marginal exactly.
    Definition in App. A.1; cells with periodogram residual above 0.02 are excluded as not spectrum-controlled. FT surrogates are used as a cross-check for the remapping artifact.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Spectrum Is Not Enough: When Context Helps Time-Series Forecasting." pith.science (2026). https://pith.science/paper/B5Y4BON4

@misc{pith2026260713006,
  author       = {Pith},
  title        = {Pith review of: The Spectrum Is Not Enough: When Context Helps Time-Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B5Y4BON4}},
  note         = {Machine review of arXiv:2607.13006}
}
abstract

A growing family of indices scores how predictable a series is from its spectrum. Practitioners increasingly read these scores as answering a different question: whether \emph{adding context}, a longer lookback, a retrieval plug-in, or a pretrained model, will help. These are not the same question. The value of context is a property of the operating point, not of the series. Any index built from the power spectrum is invariant under phase randomization, whereas the beyond-second-order value that retrieval and foundation models supply is not, because a phase-randomized series is asymptotically Gaussian. We state this as an impossibility result and isolate it with surrogate pairs that fix the spectrum and the marginal by construction. We then give a label-free, configuration-level diagnostic, the coverage deficit, whose principal term measures beyond-spectrum structure as the gain of analog over linear prediction. On seven benchmarks the prediction holds: window-keyed retrieval's value collapses across surrogate pairs (ECL median $+33\%\!\to\!-35\%$, $p{<}10^{-40}$) while every spectral index stays frozen; a foundation model's value splits into a surviving second-order part and a small beyond-linear margin that collapses; a longer linear window's value survives. Leave-one-dataset-out, the structure term predicts the sign of beyond-spectrum value where the spectral indices trail it, and the reverse holds for the second-order mechanism. We introduce no new forecaster; the contribution is the distinction, a controlled comparison, and a diagnostic for the deployment decision. Code: https://github.com/KurbanIntelligenceLab/SINE

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 8 canonical work pages

  1. [1]

    Ibram Abdelmalak, Kiran Madhusudhanan, Jungmin Choi, Christian Kl \"o tergens, Vijaya Krishna Yalavarthi, Maximilian Stubbemann, and Lars Schmidt-Thieme. Channel dependence, limited lookback windows, and the simplicity of datasets: How biased is time series forecasting? In Raymond Chi-Wing Wong, Hanghang Tong, Hua Lu, James Kwok, Flora Salim, Yuanfeng Son...

    2026

  2. [2]

    Retrieval mechanisms surpass long-context scaling in time series forecasting, 2026

    Rishi Ahuja, Kumar Prateek, Simranjit Singh, and Vijay Kumar. Retrieval mechanisms surpass long-context scaling in time series forecasting, 2026. URL https://arxiv.org/abs/2605.08217

    arXiv 2026

  3. [3]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, 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. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: Learning the language ...

    arXiv 2024

  4. [4]

    Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael Bohlke-Schneider

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael B...

    arXiv 2025

  5. [5]

    Permutation entropy: A natural complexity measure for time series

    Christoph Bandt and Bernd Pompe. Permutation entropy: A natural complexity measure for time series. Phys. Rev. Lett., 88: 0 174102, Apr 2002. doi:10.1103/PhysRevLett.88.174102. URL https://link.aps.org/doi/10.1103/PhysRevLett.88.174102

    doi:10.1103/physrevlett.88.174102 2002

  6. [6]

    There are no champions in supervised long-term time series forecasting

    Lorenzo Brigato, Rafael Morand, Knut Joar Str mmen, Maria Panagiotou, Markus Schmidt, and Stavroula Mougiakakou. There are no champions in supervised long-term time series forecasting. Transactions on Machine Learning Research, 2026. ISSN 2835-8856. URL https://openreview.net/forum?id=yO1JuBpTBB

    2026

  7. [7]

    Why do time series models need long context windows?, 2026

    Luca Butera, Giovanni De Felice, Andrea Cini, and Cesare Alippi. Why do time series models need long context windows?, 2026. URL https://arxiv.org/abs/2606.01999

    arXiv 2026

  8. [8]

    Multivariate de bruijn graphs: A symbolic graph framework for time series forecasting, 2025

    Mert Onur Cakiroglu, Idil Bilge Altun, Mehmet Dalkilic, Elham Buxton, and Hasan Kurban. Multivariate de bruijn graphs: A symbolic graph framework for time series forecasting, 2025. URL https://arxiv.org/abs/2505.22768. ICML 2025 Workshop on Foundation Models for Structured Data

    arXiv 2025

Show all 43 references
  1. [9]

    Forecastability as an information-theoretic limit on prediction, 2026

    Peter Maurice Catt. Forecastability as an information-theoretic limit on prediction, 2026. URL https://arxiv.org/abs/2603.27074

    2026

  2. [10]

    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 0 (366a): 0 427--431, 1979. doi:10.1080/01621459.1979.10482531. URL https://doi.org/10.1080/01621459.1979.10482531

    1979

  3. [11]

    Beyond model ranking: Predictability-aligned evaluation for time series forecasting, 2026

    Wanjin Feng, Yuan Yuan, Jingtao Ding, and Yong Li. Beyond model ranking: Predictability-aligned evaluation for time series forecasting, 2026. URL https://arxiv.org/abs/2509.23074

    2026 arXiv

  4. [12]

    Model-free quantification of time-series predictability

    Joshua Garland, Ryan James, and Elizabeth Bradley. Model-free quantification of time-series predictability. Phys. Rev. E, 90: 0 052910, Nov 2014. doi:10.1103/PhysRevE.90.052910. URL https://link.aps.org/doi/10.1103/PhysRevE.90.052910

    2014 doi

  5. [13]

    Forecastable Component Analysis

    Georg Goerg. Forecastable Component Analysis . In Sanjoy Dasgupta and David McAllester, editors, Proceedings of the 30th International Conference on Machine Learning, volume 28 of Proceedings of Machine Learning Research, pages 64--72, Atlanta, Georgia, USA, June 2013. PMLR. U...

    2013

  6. [14]

    Retrieval augmented time series forecasting

    Sungwon Han, Seungeon Lee, Meeyoung Cha, Sercan O Arik, and Jinsung Yoon. Retrieval augmented time series forecasting. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors, Proceedings of the 42n...

    2025

  7. [15]

    Attractor memory for long-term time series forecasting: A chaos perspective

    Jiaxi Hu, Yuehong Hu, Wei Chen, Ming Jin, Shirui Pan, Qingsong Wen, and Yuxuan Liang. Attractor memory for long-term time series forecasting: A chaos perspective. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural ...

    2024 doi

  8. [16]

    Channel-wise retrieval for multivariate time series forecasting

    Junhyeok Kang, Jun Seo, Soyeon Park, Sangjun Han, Seohui Bae, Hyeokjun Choe, and Soonyoung Lee. Channel-wise retrieval for multivariate time series forecasting. In ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1336-...

    2026

  9. [17]

    A symbolic representation of time series, with implications for streaming algorithms

    Jessica Lin, Eamonn Keogh, Stefano Lonardi, and Bill Chiu. A symbolic representation of time series, with implications for streaming algorithms. In Proceedings of the 8th ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge Discovery, DMKD '03, page 2–11, New Yo...

    2003

  10. [18]

    Cyclenet: Enhancing time series forecasting through modeling periodic patterns

    Shengsheng Lin, Weiwei Lin, Xinyi Hu, Wentai Wu, Ruichao Mo, and Haocheng Zhong. Cyclenet: Enhancing time series forecasting through modeling periodic patterns. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural In...

    2024 doi

  11. [19]

    Moirai 2.0: When less is more for time series forecasting, 2026

    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, 2026. URL https://arxiv.org/abs/2511.11698

    2026

  12. [20]

    Retrieval-augmented diffusion models for time series forecasting

    Jingwei Liu, Ling Yang, Hongyan Li, and Shenda Hong. Retrieval-augmented diffusion models for time series forecasting. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems, volume 37, p...

    2024 doi

  13. [21]

    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 B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun, editors, International Conference on Lea...

    2024

  14. [22]

    catch22: Canonical time-series characteristics selected through highly comparative time-series analysis

    Carl H Lubba, Sarab S Sethi, Philip Knaute, Simon R Schultz, Ben D Fulcher, and Nick S Jones. catch22: Canonical time-series characteristics selected through highly comparative time-series analysis. bioRxiv, 2019. doi:10.1101/532259. URL https://www.biorxiv.org/content/early/2...

    2019 doi

  15. [23]

    Lets forecast: Learning embedology for time series forecasting, 2025

    Abrar Majeedi, Viswanatha Reddy Gajjala, Satya Sai Srinath Namburi GNVV, Nada Magdi Elkordi, and Yin Li. Lets forecast: Learning embedology for time series forecasting, 2025. URL https://arxiv.org/abs/2506.06454

    2025 arXiv

  16. [24]

    Spectral retrieval-augmented time-series forecasting, 2026

    Huu Hiep Nguyen, Minh Hoang Nguyen, Dung Nguyen, and Hung Le. Spectral retrieval-augmented time-series forecasting, 2026. URL https://arxiv.org/abs/2606.19412

    2026 arXiv

  17. [25]

    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, 2023. URL https://arxiv.org/abs/2211.14730

    2023 arXiv

  18. [26]

    Nikias and M.R

    C.L. Nikias and M.R. Raghuveer. Bispectrum estimation: A digital signal processing framework. Proceedings of the IEEE, 75 0 (7): 0 869--891, 1987. doi:10.1109/PROC.1987.13824

    1987

  19. [27]

    Generating surrogate data for time series with several simultaneously measured variables

    Dean Prichard and James Theiler. Generating surrogate data for time series with several simultaneously measured variables. Phys. Rev. Lett., 73: 0 951--954, Aug 1994. doi:10.1103/PhysRevLett.73.951. URL https://link.aps.org/doi/10.1103/PhysRevLett.73.951

    1994 doi

  20. [28]

    R \"a th and R

    C. R \"a th and R. Monetti. Surrogates with Random Fourier Phases, pages 274--285. World Scientific, 2009. doi:10.1142/9789814271349_0031. URL https://www.worldscientific.com/doi/abs/10.1142/9789814271349_0031

    2009 doi

  21. [29]

    Surrogate time series

    Thomas Schreiber and Andreas Schmitz. Surrogate time series. Physica D: Nonlinear Phenomena, 142 0 (3): 0 346--382, 2000. ISSN 0167-2789. doi:https://doi.org/10.1016/S0167-2789(00)00043-9. URL https://www.sciencedirect.com/science/article/pii/S0167278900000439

    2000 doi

  22. [30]

    George Sugihara and Robert M. May. Nonlinear forecasting as a way of distinguishing chaos from measurement error in time series. Nature, 344 0 (6268): 0 734--741, Apr 1990. ISSN 1476-4687. doi:10.1038/344734a0. URL https://doi.org/10.1038/344734a0

    1990 doi

  23. [31]

    Detecting strange attractors in turbulence

    Floris Takens. Detecting strange attractors in turbulence. In Dynamical Systems and Turbulence, Warwick 1980, volume 898 of Lecture Notes in Mathematics, pages 366--381. Springer, 1981

    1980

  24. [32]

    Doyne Farmer

    James Theiler, Stephen Eubank, André Longtin, Bryan Galdrikian, and J. Doyne Farmer . Testing for nonlinearity in time series: the method of surrogate data. Physica D: Nonlinear Phenomena, 58 0 (1): 0 77--94, 1992. ISSN 0167-2789. doi:https://doi.org/10.1016/0167-2789(92)90102...

    1992

  25. [33]

    Retrieval augmented time series forecasting, 2026

    Kutay Tire, Ege Onur Taga, Muhammed Emrullah Ildiz, and Samet Oymak. Retrieval augmented time series forecasting, 2026. URL https://arxiv.org/abs/2411.08249

    2026 arXiv

  26. [34]

    Spectral predictability as a fast reliability indicator for time series forecasting model selection, 2025 a

    Oliver Wang, Pengrui Quan, Kang Yang, and Mani Srivastava. Spectral predictability as a fast reliability indicator for time series forecasting model selection, 2025 a . URL https://arxiv.org/abs/2511.08884

    2025

  27. [35]

    Time series forecastability measures, 2025 b

    Rui Wang, Steven Klee, and Alexis Roos. Time series forecastability measures, 2025 b . URL https://arxiv.org/abs/2507.13556

    2025 arXiv

  28. [36]

    Frequency matters: When time series foundation models fail under spectral shift, 2025 c

    Tianze Wang, Sofiane Ennadir, John Pertoft, Gabriela Zarzar Gandler, Lele Cao, Zineb Senane, Styliani Katsarou, Sahar Asadi, Axel Karlsson, and Oleg Smirnov. Frequency matters: When time series foundation models fail under spectral shift, 2025 c . URL https://arxiv.org/abs/2511.05619

    2025

  29. [37]

    Exploring accuracy law for deep time series forecasters: An empirical study, 2026

    Yuxuan Wang, Haixu Wu, Yuezhou Ma, Yuchen Fang, Ziyi Zhang, Yong Liu, Shiyu Wang, Zhou Ye, Yang Xiang, Jianmin Wang, and Mingsheng Long. Exploring accuracy law for deep time series forecasters: An empirical study, 2026. URL https://arxiv.org/abs/2510.02729

    2026 arXiv

  30. [38]

    On identifying why and when foundation models perform well on time-series forecasting using automated explanations and rating, 2025

    Michael Widener, Kausik Lakkaraju, John Aydin, and Biplav Srivastava. On identifying why and when foundation models perform well on time-series forecasting using automated explanations and rating, 2025. URL https://arxiv.org/abs/2508.20437

    2025 arXiv

  31. [39]

    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. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=Yd8eHMY1wz

    2024

  32. [40]

    Frequency adaptive normalization for non-stationary time series forecasting

    Weiwei Ye, Songgaojun Deng, Qiaosha Zou, and Ning Gui. Frequency adaptive normalization for non-stationary time series forecasting. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems,...

    2024 doi

  33. [41]

    Are transformers effective for time series forecasting? Proceedings of the AAAI Conference on Artificial Intelligence, 37 0 (9): 0 11121–11128, Jun

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? Proceedings of the AAAI Conference on Artificial Intelligence, 37 0 (9): 0 11121–11128, Jun. 2023. doi:10.1609/aaai.v37i9.26317. URL https://ojs.aaai.org/index.php/AAAI/art...

    2023 doi

  34. [42]

    Context parroting: A simple but tough-to-beat baseline for foundation models in scientific machine learning, 2026

    Yuanzhao Zhang and William Gilpin. Context parroting: A simple but tough-to-beat baseline for foundation models in scientific machine learning, 2026. URL https://arxiv.org/abs/2505.11349

    2026

  35. [43]

    Stationarity-aware retrieval-augmented time series forecasting, 2026

    Shiqiao Zhou, Holger Schöner, Zipeng Wu, Edouard Fouché, IAG Wilson, and Shuo Wang. Stationarity-aware retrieval-augmented time series forecasting, 2026. URL https://arxiv.org/abs/2606.04135

    2026 arXiv

Pith tools

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