Pith. sign in

REVIEW 4 major objections 3 minor 31 references

Quantum-Optimized Selective State Space Model for Efficient Time Series Prediction

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Q-SSM claims a one-qubit expectation gate stabilizes long-horizon forecasting in linear time and outperforms S-Mamba and Transformer baselines on ETT, Traffic, and Exchange.

desk verdict The paper's own equations defeat its main claim: the quantum gate is input-independent, so the model is a fixed-leak filter, not a selective SSM; the baseline table also looks copied. read the letter →

arxiv 2509.00259 v1 pith:BXHUWWTR submitted 2025-08-29 cs.LG

Stefan-Alexandru Jura , Mihai Udrescu , Alexandru Topirceanu This is my paper · ORCID
classification cs.LG
keywords timeseriesforecastinglong-termpredictionstatespacemodelsquantum-optimizedmachinelearninghybridquantum-classicalvariationalquantumcircuits
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

Q-SSM is a recurrent forecaster whose memory update is ht = (1-g)ht-1 + g*ut, with the leak coefficient g produced by a single-qubit circuit: two rotations RY(theta) and RX(phi), a Z-measurement expectation cos(theta)cos(phi), and a sigmoid clipped to [0.05, 0.95]. The paper's claim is that this quantum-produced gate stabilizes training and long-memory retention better than classical sigmoid gates, because its derivatives are bounded and the update is a contraction, while adding no quadratic attention cost. Across ETT, Traffic, and Exchange, the authors report Q-SSM winning 32 of 36 horizon-dataset cells, with 30-40% MSE reductions over Autoformer on periodic data and up to 57% on Exchange at H=96. If true, it would show that a minimal quantum component, not a large ansatz, can serve as a lightweight alternative to attention for long-horizon forecasting.

What carries the argument

The central object is the RY-RX ansatz with a Z-basis measurement. Starting from |0>, the circuit applies RY(theta) then RX(phi), and the measured observable has expectation z = <Z> = cos(theta)cos(phi) in [-1, 1]. Two such expectations are combined as s = w1*z1 + w2*z2 + bg, then g = clip(sigma(s), 0.05, 0.95). This g is the 'quantum gate' that interpolates the state-space update ht = (1-g)ht-1 + g*ut. Its load-bearing properties are: (i) the derivatives dz/dtheta = -sin(theta)cos(phi) and dz/dphi = -cos(theta)sin(phi) are each bounded by 1, making the whole gate 1-Lipschitz per parameter when combined with sigma' <= 1/4; and (ii) clipping keeps g inside (0,1), so ||dht/dht-1|| = |1-g| < 1,

What would settle it

On a fixed test set (e.g., ETTm1, H=96), take the trained Q-SSM and replace the two cosine expectations with the constants 1 and 0, then re-optimize only w1, w2, bg so g is a single trainable leak; if MSE/MAE remain within run-to-run noise of the full model, the quantum circuit's parameter geometry is not what produces the gains. A second decisive check: rerun S-Mamba and Q-SSM from the same random seeds on the exact same 60/20/20 split and report error bars; if the reported margins of about 2% swap signs under reseeding, the claimed state-of-the-art status rests on borrowed baseline numbers.

Watch

Extended reading notes

Core claim

On its own terms, Q-SSM's discovery is that a minimal variational quantum gate can replace the classical pre-activation of an SSM's update gate and thereby stabilize long-horizon forecasting without quadratic attention. The core recurrence is ht = (1-g)ht-1 + g*LN(W(P(xt)) + b + alpha*c); the gate is g = clip(sigma(w1 cos(theta1)cos(phi1) + w2 cos(theta2)cos(phi2) + bg), 0.05, 0.95), where each cos(theta)cos(phi) is the Z-expectation of a one-qubit circuit |psi> = RX(phi)RY(theta)|0>. Because the quantum expectation has derivatives bounded by 1 and sigma' <= 1/4, the gate is Lipschitz in its parameters, and because g is clipped away from 0 and 1, the recurrence's Jacobian has norm |1-g| < 1,

Load-bearing premise

The load-bearing premise is that a gate value computed only from learned parameters, never from the current input x_t, still counts as an 'adaptive' or 'selective' memory controller; at inference g is the same scalar for every time step and every feature, so the recurrence is a fixed-leak linear filter, and the claimed quantum-adaptive memory control and the comparison to S-Mamba's selective updates rest on that premise.

Editorial extensions

If this is right

  • A fixed backbone width (k=d=128) with the same hyperparameters transfers across six benchmarks, so Q-SSM implies that minimal recurrent capacity plus a bounded quantum gate is enough to compete with much larger attention stacks.
  • Because the recurrence Jacobian is bounded by 1-g < 1, Q-SSM implies that hidden state information is neither amplified nor erased over thousands of steps, which is exactly what makes 720-step forecasts stable.
  • The quantum gate's cost is two single-qubit expectation values per step, so Q-SSM preserves linear-time inference O(W(Fk+kd)+HdF), avoiding the O(W log W) of Autoformer and the O(Wd^2) hidden-dimension cost of Mamba.
  • On the normalized benchmarks, Q-SSM reports 32/36 best MSE/MAE cells, with the residual decoder predicting deviations from the last observed value rather than absolute values; this supports the claim that grounding forecasts in the last observation prevents long-horizon drift.

Reading between the lines

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

  • Because g in Eq. (15) does not depend on x_t, a straightforward test is to feed the same gate parameters different input windows; the model's per-step leakage is unchanged, so the paper's 'adaptive' language is currently a training-time property only. Making theta and phi functions of the input would turn the fixed-leak filter into a genuinely selective recurrence, and the reported margins over S-
  • The bounded oscillatory pre-activation cos(theta)cos(phi) is classical; replacing the two single-qubit expectations with a classical sinusoidal nonlinearity of the same Lipschitz constant would isolate whether the benefit is quantum-specific or simply the bounded, non-saturating gating geometry.
  • At the longest horizons on ETTh2 and Traffic, S-Mamba outperforms Q-SSM by about 2%; a split-matched head-to-head with multiple seeds would show whether the fixed-leak recurrence hits a horizon-dependent ceiling that an input-conditional gate could lift.
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 / 3 minor

Summary. The manuscript proposes Q-SSM, a hybrid quantum-classical recurrent model for long-horizon multivariate time series forecasting. The architecture combines a linear-time gated state-space recurrence with a gate produced by a single-qubit variational circuit (RY-RX rotations and a Pauli-Z expectation), followed by a sigmoid and clipping. The authors claim that this quantum gate adaptively regulates memory updates, improves stability, and yields state-of-the-art accuracy on ETT, Traffic, and Exchange benchmarks relative to Transformer, RNN, TCN, and S-Mamba baselines.

Significance. If the central claims were true, the paper would make a useful contribution: an O(T)-complexity forecasting model with a simple, classically simulable quantum gate improving over strong baselines would be of interest to both the time-series and quantum-machine-learning communities. The paper also offers explicit formulas for the gate, Lipschitz bounds, and contraction of the recurrence, which are transparent. However, the central architectural claim is undermined by the gate's lack of input dependence, and the empirical evaluation is compromised by apparent data errors and an incompatible evaluation protocol. The positive aspects are the explicit mathematical presentation and the reproducibility attempt via a code link, but these do not offset the load-bearing flaws.

major comments (4)
  1. [Sec. III-B, III-C, Eqs. (9), (12)-(15)] The gate g as defined in Eq. (15) has no dependence on the input x_t, the hidden state h_{t-1}, or any time-varying quantity. From Eqs. (12)-(15), s = w1 cos(theta1) cos(phi1) + w2 cos(theta2) cos(phi2) + b_g, and g = clip(sigma(s), gmin, gmax), so after training g is a single constant scalar applied at every time step and every feature. The recurrence (9) is therefore a linear time-invariant fixed-leak filter, not a selective state space model in the sense of Mamba/S-Mamba, where selection requires input-dependent B, C, and Delta. The claimed 'adaptive quantum memory control' and the model's central selective-SSM framing collapse. The stability bounds in Eqs. (16)-(21) are mathematically correct but only show that a parameter-dependent gate is Lipschitz and contractive; they do not establish input adaptivity.
  2. [Table I] The baseline results contain identical rows across different datasets, which is not plausible. For example, Informer reports 0.365/0.453 at H=96 for both ETTm2 and ETTh1, and the entire block of LogTrans, Reformer, LSTNet, LSTM, and TCN values is identical for many rows across ETT variants. Since these datasets have different statistics and lengths, such duplication indicates a data-handling error or copying of published tables without corresponding re-runs. This undermines the empirical comparison and the claim that Q-SSM 'consistently' outperforms baselines.
  3. [Sec. IV-B vs. Sec. V] The evaluation protocol uses a 60/20/20 chronological split, whereas the standard splits in the long-horizon forecasting literature (e.g., for ETT, Traffic, and Exchange) are typically different, often 70/20/10 or 12/4/4 months. Comparing Q-SSM results obtained on one split with published baseline numbers obtained on another split is methodologically invalid. The 'wins' and relative improvement percentages in Section V are therefore not strictly interpretable. The authors should re-run all baselines under the same data split, or clearly justify why the published numbers remain comparable.
  4. [Sec. V-g (Overall assessment)] The paper states that Q-SSM achieves 'statistically significant improvements over established baselines,' but no error bars, confidence intervals, significance tests, or repeated-seed experiments are reported. Several claimed wins over S-Mamba are extremely small (e.g., ETTm1 H=96: 0.330 vs 0.333; ETTh2 H=192: 0.372 vs 0.376), well within typical run-to-run variation. Without statistical evidence, the significance claim is unsupported.
minor comments (3)
  1. [References] References [11] and [19] are duplicates of the same S4 paper (Gu, Goel, and Re, ICLR 2021). One should be removed or replaced with the appropriate distinct citation.
  2. [Fig. 1 caption] The description of the classical wire c and 'measurement outcome stored in a one-bit classical register' is confusing because the paper explicitly uses the expectation value <Z>, not the discrete outcome. This could be clarified.
  3. [Sec. III-C4] The paragraph argues that the quantum pre-activation creates 'a richer family of input distributions for the sigmoid.' Since the pre-activation s in Eq. (14) contains no input, there are no input distributions; the wording should be corrected or the gate must be redesigned.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: Q-SSM's gate is fixed at inference, which undermines the 'selective' claim but is not a circularity.

full rationale

I examined the claimed derivation chain: the quantum gate g is defined by Eqs. (12)-(15) as a bounded sigmoid of a constant combination of trainable parameters, and the stability/contractivity bounds in Eqs. (16)-(21) follow directly from those explicit definitions. No fitted constant is renamed as a prediction; the empirical evaluation is checked against external baselines; and the reference to the authors' own earlier work [20] is not load-bearing. The central weakness—that g has no dependence on the input x_t, making the recurrence a fixed-leak linear filter rather than an input-adaptive selective SSM—is a substantive architectural validity and interpretation concern, but it is not a circularity of the derivation. Similarly, the duplicated baseline rows and missing significance tests are reproducibility/correctness concerns, not evidence that the paper's conclusions are equivalent to its inputs by construction. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The paper depends on standard single-qubit math and on several empirical choices. The gate parameters and clipping bounds are free parameters; the calendar features and split comparability are domain assumptions. No genuinely new physical entity is introduced; the 'quantum gate' is a classically simulated trigonometric function.

free parameters (5)
  • RY-RX rotation angles (theta1, phi1, theta2, phi2)
    Learned via backpropagation/parameter-shift (Eq. 6); these define the gate and are the paper's claimed contribution. See Eqs. (12)-(15).
  • Gate combination weights and bias (w1, w2, b_g)
    Learnable linear weights that map the two expectation values into the sigmoid pre-activation s.
  • Clipping bounds (gmin, gmax) = 0.05, 0.95
    Chosen empirically in Section III-C to keep the gate in a safe range; they are not derived from data.
  • Calendar influence scalar alpha
    Learnable scalar controlling the contribution of calendar features in Eq. (9).
  • Training hyperparameters (Adam lr=1e-3, weight decay=1e-4, early stopping patience=10, dropout=0.1, k=d=128) = as listed
    Chosen by hand and fixed across datasets in Section IV-A; required to reproduce the reported results.
assumptions (5)
  • standard math Single-qubit expectation value: <Z> = cos(theta)cos(phi) for RX(phi)RY(theta)|0>
    Standard quantum mechanics for the RY-RX circuit; used in Eq. (13).
  • standard math Parameter-shift rule (Eq. 6) gives exact gradients for single-qubit expectations
    Standard result in variational quantum algorithms; enables end-to-end training on the simulator.
  • domain assumption The 60/20/20 chronological split yields numbers comparable to published baseline numbers
    Section IV-B sets the split; cited baselines in Table I were not rerun, and many prior long-horizon papers use a 70/10/20 split, so this assumption is load-bearing for the empirical wins.
  • domain assumption Calendar sin/cos encodings capture relevant periodic structure for ETT and Traffic
    Section IV-A/B relies on these features to explain strong results on periodic benchmarks; without them performance may degrade.
  • ad hoc to paper Classical simulation of a one-qubit expectation value is meaningfully 'quantum' for forecasting
    The entire quantum framing rests on calling cos(theta)cos(phi) a quantum gate; no multi-qubit entanglement or hardware is used (Section III-C).
invented entities (1)
  • Variational quantum gate (two single-qubit RY-RX circuits)
    purpose: Serves as the memory update gate g in the recurrence, Eqs. (9)-(10).
    It is a classically simulated trigonometric function with no external falsifiable handle. The gate is input-independent at inference, so it is not a new physical entity and provides no evidence outside the paper's own model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum-Optimized Selective State Space Model for Efficient Time Series Prediction." pith.science (2026). https://pith.science/paper/BXHUWWTR

@misc{pith2026250900259,
  author       = {Pith},
  title        = {Pith review of: Quantum-Optimized Selective State Space Model for Efficient Time Series Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BXHUWWTR}},
  note         = {Machine review of arXiv:2509.00259}
}
read the original abstract

Long-range time series forecasting remains challenging, as it requires capturing non-stationary and multi-scale temporal dependencies while maintaining noise robustness, efficiency, and stability. Transformer-based architectures such as Autoformer and Informer improve generalization but suffer from quadratic complexity and degraded performance on very long time horizons. State space models, notably S-Mamba, provide linear-time updates but often face unstable training dynamics, sensitivity to initialization, and limited robustness for multivariate forecasting. To address such challenges, we propose the Quantum-Optimized Selective State Space Model (Q-SSM), a hybrid quantum-optimized approach that integrates state space dynamics with a variational quantum gate. Instead of relying on expensive attention mechanisms, Q-SSM employs a simple parametrized quantum circuit (RY-RX ansatz) whose expectation values regulate memory updates adaptively. This quantum gating mechanism improves convergence stability, enhances the modeling of long-term dependencies, and provides a lightweight alternative to attention. We empirically validate Q-SSM on three widely used benchmarks, i.e., ETT, Traffic, and Exchange Rate. Results show that Q-SSM consistently improves over strong baselines (LSTM, TCN, Reformer), Transformer-based models, and S-Mamba. These findings demonstrate that variational quantum gating can address current limitations in long-range forecasting, leading to accurate and robust multivariate predictions.

Figures

Figures reproduced from arXiv: 2509.00259 by the authors.

Figure 1
Figure 1. Quantum gating circuit: the qubit, initialized in [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 3
Figure 3. Decoder with MLP and residual connection. The MLP maps [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 2
Figure 2. Encoder block with quantum gating. The input [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages

  1. [1]

    D. C. Montgomery, C. L. Jennings, and M. Kulahci, Introduction to time series analysis and forecasting . John Wiley & Sons, 2015

    work page 2015

  2. [2]

    G. E. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung, Time Series Analysis: Forecasting and Control, 5th ed. John Wiley & Sons, 2015

    work page 2015

  3. [3]

    G. E. Box and G. M. Jenkins, Time Series Analysis: Forecasting and Control. Holden-Day, 1970

    work page 1970

  4. [4]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  5. [5]

    Learning phrase representations using rnn encoder–decoder for statistical machine translation,

    K. Cho, B. van Merri ¨enboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representations using rnn encoder–decoder for statistical machine translation,” in EMNLP, 2014, pp. 1724–1734

    work page 2014

  6. [6]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” arXiv preprint arXiv:1803.01271, 2018

    arXiv 2018

  7. [7]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , vol. 30, 2017

    work page 2017

  8. [8]

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

    H. Wu, J. Xu, J. Wang, and M. Long, “Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting,” in NeurIPS, 2021, pp. 22 419–22 430

    work page 2021

Show all 31 references
  1. [9]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in AAAI, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  2. [10]

    Fedformer: Frequency enhanced decomposed transformer for long-term series fore- casting,

    Y . Zhou, Z. Ma, Q. Wen, L. Sun, R. Jin, and M. Long, “Fedformer: Frequency enhanced decomposed transformer for long-term series fore- casting,” in ICML, 2022, pp. 27 268–27 286

    2022

  3. [11]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,” in ICLR, 2022

    2022

  4. [12]

    Is mamba effective for time series forecasting?

    Z. Wang, F. Kong, S. Feng, M. Wang, X. Yang, H. Zhao, D. Wang, and Y . Zhang, “Is mamba effective for time series forecasting?” arXiv preprint arXiv:2403.11144, 2024

    2024 arXiv

  5. [13]

    Reformer: The efficient transformer,

    N. Kitaev, L. Kaiser, and A. Levskaya, “Reformer: The efficient transformer,” in International Conference on Learning Representations (ICLR), 2020. [Online]. Available: https://openreview.net/forum?id= rkgNKkHtvB

    2020

  6. [14]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  7. [15]

    Macroeconomics and reality,

    C. A. Sims, “Macroeconomics and reality,” Econometrica, vol. 48, no. 1, pp. 1–48, 1980

    1980

  8. [16]

    R. J. Hyndman, A. B. Koehler, J. K. Ord, and R. D. Snyder, Forecasting with Exponential Smoothing: The State Space Approach . Springer, 2008

    2008

  9. [17]

    Learning repre- sentations by back-propagating errors,

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning repre- sentations by back-propagating errors,” Nature, vol. 323, no. 6088, pp. 533–536, 1986

    1986

  10. [18]

    Time-series forecasting with deep learning: a survey,

    B. Lim and S. Zohren, “Time-series forecasting with deep learning: a survey,” Philosophical Transactions of the Royal Society A , vol. 379, no. 2194, p. 20200209, 2021

    2021

  11. [19]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,” in International Conference on Learning Representations, 2021

    2021

  12. [20]

    Design for dependability in emerging technologies,

    L. Prodan, M. Udrescu, O. Boncalo, and M. Vladutiu, “Design for dependability in emerging technologies,” J. Emerg. Technol. Comput. Syst. , vol. 3, no. 2, p. 6–es, Jul. 2007. [Online]. Available: https://doi.org/10.1145/1265949.1265952

    2007

  13. [21]

    M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information: 10th Anniversary Edition . Cambridge, UK: Cambridge University Press, 2010

    2010

  14. [22]

    Lecture notes for physics 229: Quantum information and computation,

    J. Preskill, “Lecture notes for physics 229: Quantum information and computation,” California Institute of Technology, 1998. [Online]. Available: http://theory.caltech.edu/∼preskill/ph229/

    1998

  15. [23]

    Quantum machine learning in feature hilbert spaces,

    M. Schuld and N. Killoran, “Quantum machine learning in feature hilbert spaces,” Physical review letters, vol. 122, no. 4, p. 040504, 2019

    2019

  16. [24]

    Quantum machine learning,

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671, pp. 195–202, 2017

    2017

  17. [25]

    Circuit-centric quantum classifiers,

    M. Schuld, A. Bocharov, K. Svore, and N. Wiebe, “Circuit-centric quantum classifiers,” Physical Review A , vol. 101, no. 3, p. 032308, 2020

    2020

  18. [26]

    Variational quantum algorithms,

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, “Variational quantum algorithms,” Nature Reviews Physics, vol. 3, no. 9, pp. 625–644, 2021

    2021

  19. [27]

    Long-term time series forecasting based on decomposition and neural ODEs,

    S. Lim, Y . Nam, Y . Kee, and J. Lee, “Long-term time series forecasting based on decomposition and neural ODEs,” in 2023 IEEE International Conference on Big Data (BigData) , 2023, pp. 4614–4623

    2023

  20. [28]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  21. [29]

    Etdataset: Electricity transformer dataset,

    H. Zhou, “Etdataset: Electricity transformer dataset,” https://github.com/ zhouhaoyi/ETDataset, 2021, accessed 26 Aug 2025

    2021

  22. [30]

    Modeling long- and short- term temporal patterns with deep neural networks,

    G. Lai, W.-C. Chang, Y . Yang, and H. Liu, “Modeling long- and short- term temporal patterns with deep neural networks,” in Proceedings of SIGIR, 2018

    2018

  23. [31]

    Multivariate time series data repository,

    G. Lai, “Multivariate time series data repository,” https://github.com/ laiguokun/multivariate-time-series-data, 2018, accessed 26 Aug 2025

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.