Pith. sign in

REVIEW 3 major objections 7 minor 49 references

Channel Normalization for Time Series Channel Identification

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Channel Normalization — per-channel scale and shift parameters inside the encoder — restores channel identifiability and cuts average squared forecast error by 4% to 12% across four backbones and twelve datasets.

desk verdict A simple, well-tested normalization trick with one overclaimed extension; worth a serious referee. read the letter →

arxiv 2506.00432 v1 pith:D6H6IJK2 submitted 2025-05-31 cs.LG cs.AIstat.ML

Seunghan Lee , Taeyoung Park , Kibok Lee This is my paper
classification cs.LGcs.AIstat.ML MSC 68T0762M10
keywords channelidentifiabilitytimeseriesforecastingnormalizationlayerchannel-specificparametersfoundationmodelsinformationtheoryentropyanalysis
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

This paper argues that many time series forecasting models are channel-blind: when two channels receive identical inputs, the model produces identical outputs, ignoring which sensor or series each channel represents. It proposes Channel Normalization (CN), a drop-in replacement for Layer Normalization that gives each channel its own scale and shift parameters, so the model can tell channels apart at negligible extra cost. Two extensions follow: Adaptive CN (ACN) mixes per-channel parameters using input-dependent channel similarities, and Prototypical CN (PCN) attaches parameters to learnable prototypes instead of channels, letting one foundation model serve datasets with different or unseen channel counts. Across four backbones and twelve datasets, CN and ACN reduce average MSE by 4.0% to 12.4%, with the largest gains for channel-blind models, and PCN improves a time series foundation model on forecasting, classification, and zero-shot tasks. The paper also gives an information-theoretic account: CN raises the entropy of feature and channel representations, and datasets with larger entropy gains show larger accuracy gains.

What carries the argument

The carrying mechanism is the affine normalization step inside the model's encoder. Layer Normalization applies one learned scale $\alpha$ and shift $\beta$ to every channel, while CN keeps $C$ separate pairs $(\alpha_c, \beta_c) \in \mathbb{R}^D$, so the normalization layer itself encodes channel identity. ACN builds on this by computing a channel similarity matrix — cosine similarity between channel representations passed through a softmax with temperature $\tau$ — and using it to form dynamic local parameters $\hat{\alpha}^L_{b,c} = \sum_i \hat{S}_{b,c,i} \alpha^L_i$, combined element-wise with the global per-channel parameters. PCN replaces per-channel parameters with $K$ learnable prototypes $\alpha^P_k \in \mathbb{R}^D$, attended to by a channel-prototype similarity computed in data space through a projection layer, so an arbitrary number of channels can share the same $K$ parameters. The supporting analysis is Gaussian entropy, $H(Z) = \frac{1}{2}\log\big((2\pi e)^D \det(\Sigma)\big)$ estimated on feature and channel covariance matrices, used to argue that CN monotonically increases the joint entropy relative to layer normalization and that higher conditional entropy of representations can lower achievable forecast error.

What would settle it

A direct control experiment would compare ACN against a version that uses the same softmax-weighted parameter mixtures but with fixed random weights or similarity computed on untrained representations; if the input-dependent mixture performs no better than the random one on data-independent backbones such as RMLP and TSMixer, then the similarity mechanism is not what carries ACN's gains. A second check: build a synthetic dataset with two channels holding identical value sequences but distinct semantic identities, and verify that a non-CID model plus CN actually produces different per-channel forecasts with lower error than Layer Normalization; if the outputs remain identical, the claimed CID mechanism fails outright.

Watch

Extended reading notes

Core claim

The paper's central claim is that channel identifiability (CID) — a model's ability to produce different outputs for different channels even when their input values coincide — is a real and neglected failure mode of time series models, and that the fix belongs in the normalization layer. A model counts as channel-identifiable if there exists an input with two identical channel series for which the outputs differ; Layer Normalization, which applies one shared affine transformation across channels, actively suppresses this ability. Channel Normalization replaces the shared $\{\alpha, \beta\}$ with per-channel $\{\alpha_c, \beta_c\}$, so that $\hat{z}_{b,c,d} = \alpha_c \cdot \mathrm{Norm}(z_{b,c,d}) + \beta_c$, and this single change improves forecasting on every dataset and backbone tested, with the largest gains on non-CID models such as iTransformer and RMLP. Adaptive CN adds input-dependent parameter mixing and helps most for data-independent models such as RMLP and TSMixer, while Prototypical CN extends the idea to unknown channel counts by parameterizing prototypes rather than channels. The paper reports average MSE reductions of 4.0% to 12.4% depending on the backbone, and its Gaussian entropy analysis shows that CN enriches feature representations, increases the uniqueness of channel representations, and diversifies attention heads and inter-channel correlations, with a 0.724 correlation between channel-entropy gain and MSE improvement across datasets.

Load-bearing premise

The load-bearing premise is that cosine similarity between channel representations, passed through a softmax, is a trustworthy signal for mixing per-channel normalization parameters in ACN; this is a heuristic rather than a derived principle, and on two settings in the paper's own tables (S-Mamba on PEMS04 and Weather) ACN performs slightly worse than plain CN, so if the similarity signal is noisy the claimed extra benefit of ACN over CN would not hold generally.

Editorial extensions

If this is right

  • Replacing Layer Normalization with Channel Normalization inside the encoder improves forecasting on every one of the 12 tested datasets across all four backbones, with average MSE reductions of 11.3% for iTransformer, 5.7% for RMLP, 4.0% for S-Mamba, and 4.5% for TSMixer.
  • The extra gain from Adaptive CN tracks the backbone's data dependency: data-independent models such as RMLP and TSMixer benefit more from input-adaptive parameters than data-dependent models such as iTransformer and S-Mamba.
  • Prototypical CN lets a single time series foundation model train on datasets with varying channel counts and forecast on unseen datasets; on UniTS it improves performance across 20 forecasting and 18 classification tasks in both supervised and prompt-tuning settings, and improves zero-shot forecasting on unseen datasets and horizons.
  • Channel entropy gain tracks benefit: across datasets, larger entropy gains from CN correlate with larger MSE improvements (correlation 0.724), consistent with the information-theoretic account of why CN works.
  • CN is cheap and composable: it is orthogonal to input-level normalization such as RevIN, and on iTransformer with PEMS08 it adds about 0.7M parameters and minimal runtime while outperforming channel identifiers and C-LoRA.

Reading between the lines

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

  • Because the mechanism is effectively a soft clustering of channels, a natural untested extension is to lift ACN's softmax-weighted parameter mixture out of the normalization layer and use it as a general channel-attention module; the paper's framing predicts that any architecture that mixes or pools channels — attention over channel tokens, bidirectional state-space scans — risks channel-identifia
  • The entropy diagnostic suggests a cheap selection tool: estimate channel entropy gain after a short training run to predict which datasets and backbones will benefit most from channel-identifiability enhancement, instead of running full forecasting grids.
  • PCN's prototypes act as a learned cluster book, and the paper's own distortion plots show redundant prototypes beyond K = 20; an automatic knee-point selection for K appears achievable and would remove the remaining hyperparameter.
  • The paper itself reports a convergence issue in the UniTS prompt-tuning setting and lowers the hidden dimension uniformly for both baseline and PCN to keep the comparison fair; this preserves validity, but it means the prompt-tuning results are measured below the originally published UniTS performance and should be read as relative comparisons.
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 / 7 minor

Summary. The paper proposes Channel Normalization (CN), a normalization layer that replaces the shared affine parameters of Layer Normalization with per-channel affine parameters, thereby giving time-series models the ability to distinguish channels even when input values are identical. Two extensions are proposed: Adaptive CN (ACN), which mixes per-channel local parameters with weights derived from input cosine similarities, and Prototypical CN (PCN), which assigns parameters to learned prototypes so that models can handle datasets with unknown or varying channel counts. The authors evaluate CN/ACN on four backbones and twelve datasets and PCN on the UniTS foundation model, reporting average MSE improvements of 4.0%–12.4%. They also offer an information-theoretic analysis claiming that CN enriches feature representations, increases channel uniqueness, and diversifies attention heads and channel correlations.

Significance. If the central claims hold, CN is a simple, architecture-agnostic plug-in that improves forecasting across both channel-identifiable and channel-unidentifiable models, and PCN is a plausible mechanism for extending CID-style normalization to foundation models. The paper's strengths include the breadth of the empirical evaluation (4 backbones × 12 datasets × 4 horizons), the isolation of the mechanism through comparisons with C-token, C-project, channel-identifier, and C-LoRA, the ablation of ACN's global/local parameters, and extensive robustness checks on the temperature, prototype count, and similarity space. The code is publicly released. However, the claim that ACN yields additional gains over CN is contradicted by several entries in the paper's own Table 2, no error bars or significance tests are provided, and the theoretical entropy justification in Appendix C conflates the entropy of a representation with the entropy of a tuple containing the representation and its normalization parameters. These issues affect the strength and interpretation of two of the three headline contributions, though the core CN idea appears defensible.

major comments (3)
  1. [§5.1, Table 2] The statement that "both CN and ACN consistently improve across all datasets and backbones, with ACN yielding additional gains compared to CN" is not supported by Table 2. Counterexamples include S-Mamba on PEMS04 (CN MSE 0.085 vs. ACN 0.095), S-Mamba on Weather (0.246 vs. 0.247), TSMixer on ETTh1 (0.438 vs. 0.453), and TSMixer on PEMS08 MAE (0.250 vs. 0.258). Combined with the complete absence of error bars, multiple random seeds, or significance tests, the average ACN-vs-CN advantage (1.1–6.1 percentage points) cannot be distinguished from noise, and it is driven by a small number of large PEMS gains while regressions occur on other settings. The ACN extension should either be reported as beneficial on a subset of settings, or the claims should be accompanied by statistical evidence and a discussion of when the cosine-similarity mixing of Equations (4)–(6) helps or hurts.
  2. [Appendix C, Eqs. (C.2)–(C.5) and (C.9)–(C.11)] The theoretical entropy analysis is not valid as written. In Eqs. (C.2)–(C.5), the proof derives H(Z) ≤ H(Z, α1, β1) ≤ H(Z, {αi, βi}i=1^C), but the last quantity is the joint entropy of the representation and the normalization parameters, not the entropy of the post-CN representation Z_CN. The non-negativity of conditional entropy only shows that appending parameters to a tuple increases the tuple's entropy; it does not show that the normalized representation itself has higher marginal entropy. Similarly, Justification 2 states that "a more informative representation (i.e., higher H(Z)) can potentially lower forecasting error" and invokes the MMSE bound, but the bound is in terms of H(Y|Z), and a higher marginal entropy H(Z) does not imply a lower conditional entropy H(Y|Z); a noisier representation can have higher entropy while being less predictive. The empirical correlation in Figure 4(b) (ρ = 0.724) is a post-hoc correlation between entropy gain and MSE improvement and does not establish a causal or information-theoretic link. These arguments should be recast as heuristic observations or corrected to state the actual conditions under which the bound applies.
  3. [§5.2, Table 4 and Appendix N] The claim of "consistent improvements observed across all tasks" for PCN on UniTS is contradicted by the appendix results. In Table N.2 the supervised setting wins only 16 of 20 forecasting tasks, and in Table N.3 the supervised classification setting wins 16 of 18, with regressions such as JapaneseVowels (93.5→92.7), NonInvasive (90.5→89.7), and prompt-tuning PEMS-SF (85.0→82.7). The aggregate improvements are real (e.g., average forecast MSE 0.469→0.433), but the text should report the distribution of per-task wins/losses and note the regressions, especially because the convergence issue described in Appendix A.2 may affect the prompt-tuning baseline. This is important for calibrating the PCN claim, which is one of the paper's three headline contributions.
minor comments (7)
  1. [General] No standard errors, confidence intervals, or significance tests are reported for any table, including the small-gain entries (e.g., RMLP on ETTh2 and Exchange in Table 2). At minimum, the paper should report the number of runs and the variance for the headline averages.
  2. [Table 2 / Table M.1] The rows labeled "Best count (/48)" and "1stCount"/"2ndCount" are not defined in the text or captions; please explain exactly what these counts represent and how ties are handled.
  3. [Appendix G] The heading "Comparison with contant vectors" contains a typo; it should read "constant vectors."
  4. [Figure 4(b)] The reported correlation ρ = 0.724 should be accompanied by the number of points, a confidence interval, and the identity of the datasets; with only 12 datasets, the correlation is not very informative.
  5. [§6.3, Figure 8] The t-SNE visualization is qualitative; the claim that "four distinct clusters" are formed and that cluster 4 lacks close relationships should be supported by a quantitative cluster-quality measure (e.g., silhouette score) or by the correlation values shown in panel (c).
  6. [Appendix A.2] The sentence about a convergence issue in prompt-tuning references a GitHub issue but provides no link or issue number; please add the reference so readers can reproduce the setting.
  7. [§4.2, Eq. (6)] The notation αL_c is used both for the static local parameter and, via the weighted average in Eq. (6), for the dynamic parameter α̂L_{b,c}; please distinguish the two consistently, e.g., by naming the static parameter θL_c and the output φL_{b,c}.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity: the entropy 'justification' is a definitional tautology; the empirical central claims are benchmarked and self-contained.

  1. self definitional [Appendix C, Theoretical Entropy Analysis, Justification 1, Eqs. (C.1)-(C.5)]
    "Justification 1. Applying CN achieves a more informative representation (ZCN) compared to LN (ZLN) or without any normalization (ZNone), as it increases in the entropy : H(ZNone) ≤ H(ZLN) ≤ H(ZCN). Proof. The joint entropy can be decomposed as follows: H(Z) =HNone ≤ H(Z) + H(α1, β1|Z) = H(Z, α1, β1) =HLN ≤ H(Z, α1, β1) + H({αi, βi}C i=2|α1, β1) = H(Z, {αi, βi}C i=1) =HCN. This follows from the non-negativity of conditional entropy."

    The proof constructs ZLN := (Z, α1, β1) and ZCN := (Z, {α_i, β_i}), so the inequality H(ZCN) ≥ H(ZLN) follows from the non-negativity of conditional entropy for arbitrary added random variables; it does not use the normalization operation, the learned parameter values, or any property of CN beyond relabeling the augmented random vector. The claimed 'more informative representation' is therefore the definitional inequality H(Z, more parameters) ≥ H(Z, fewer parameters), presented as a theoretical derivation of CN's benefit. This step is not load-bearing for the empirical forecasting results, which are benchmarked against external baselines, but it is offered as a theoretical justification and reduces by construction to the definition of joint entropy.

full rationale

The paper's central claims are empirical: CN/ACN/PCN improve forecasting MSE/MAE across four backbones and twelve datasets, with comparisons against external baselines (LN, channel identifier, C-LoRA, IN, UniTS) and ablations over K, τ, and similarity spaces. These results are not fitted to a predefined outcome; hyperparameters are selected by validation loss, and the reported gains are externally benchmarked. The ACN-versus-CN contradictions in Table 2 (e.g., S-Mamba PEMS04 CN 0.085 vs ACN 0.095) are correctness risks, not circularity. The entropy-gain correlation (ρ = 0.724) is post-hoc and descriptive, not a predictive fit. The only circular step found is the 'theoretical entropy' Justification 1, where the entropy ordering is obtained by defining the representation to include the affine parameters, so the conclusion is a tautology of joint entropy rather than a property of CN. Because this tautology appears in an explanatory analysis and does not support the main benchmarking claims, the overall circularity is minor: score 2.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method introduces additional learned parameters at each normalization layer (2*C*D for CN), but these are standard model parameters, not external free constants. The key hyperparameters are K (number of prototypes) and tau (softmax temperature). The theoretical analysis relies on a Gaussian assumption that is not validated for time series data.

free parameters (4)
  • CN per-channel scale and shift parameters (alpha_c, beta_c) = learned per channel
    Core mechanism of the method; these affine parameters are fitted to the training data at each normalization layer.
  • ACN global and local parameters (alpha_G, alpha_L, beta_G, beta_L) = learned
    Additional parameters for the adaptive variant; the local parameters are combined via a similarity-weighted sum.
  • PCN prototype parameters (alpha_P, beta_P), K x D = learned, K chosen as a hyperparameter
    Prototype vectors assigned to channels via a softmax similarity; K is set to 5 for single-task models and varied in ablations for TSFMs.
  • Softmax temperature tau = default 0.5
    Temperature in the similarity softmax for ACN and PCN; the paper shows robustness across 0.05 to 1.0, so it is not a critical fitted constant.
assumptions (3)
  • domain assumption Backbone architectures (iTransformer, RMLP, S-Mamba, TSMixer, UniTS) are accepted as given and correctly implemented.
    The method is evaluated on top of these models; any bug or misconfiguration could affect results. The paper states that baselines are obtained from official code.
  • ad hoc to paper Gaussian entropy and MMSE bounds apply to time series representations.
    Appendix C assumes multivariate Gaussian distributions to derive entropy estimates and the MMSE bound; time series features are not shown to be Gaussian, so the theoretical justification is not generally applicable.
  • domain assumption Standard chronological train/validation/test splits and hyperparameter tuning are sufficient for generalization claims.
    No error bars or multiple seeds are reported; the paper relies on the standard practice in the field, which limits the statistical strength of the conclusions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Channel Normalization for Time Series Channel Identification." pith.science (2026). https://pith.science/paper/D6H6IJK2

@misc{pith2026250600432,
  author       = {Pith},
  title        = {Pith review of: Channel Normalization for Time Series Channel Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D6H6IJK2}},
  note         = {Machine review of arXiv:2506.00432}
}
read the original abstract

Channel identifiability (CID) refers to the ability to distinguish between individual channels in time series (TS) modeling. The absence of CID often results in producing identical outputs for identical inputs, disregarding channel-specific characteristics. In this paper, we highlight the importance of CID and propose Channel Normalization (CN), a simple yet effective normalization strategy that enhances CID by assigning distinct affine transformation parameters to each channel. We further extend CN in two ways: 1) Adaptive CN (ACN) dynamically adjusts parameters based on the input TS, improving adaptability in TS models, and 2) Prototypical CN (PCN) introduces a set of learnable prototypes instead of per-channel parameters, enabling applicability to datasets with unknown or varying number of channels and facilitating use in TS foundation models. We demonstrate the effectiveness of CN and its variants by applying them to various TS models, achieving significant performance gains for both non-CID and CID models. In addition, we analyze the success of our approach from an information theory perspective. Code is available at https://github.com/seunghan96/CN.

Figures

Figures reproduced from arXiv: 2506.00432 by the authors.

Figure 1
Figure 1. Motivating example for channel identifiability. When two different channels receive the locally identical inputs (green), a non-CID model yields the same outputs (yellow) for both, failing to distinguish between them, as shown in the left panel. In contrast, applying CN enables CID and produces distinct outputs even with the same inputs, as shown in the right panel. Average MSE (4Hs) ETTm1 Weather PEMS03 Imp. iTrans… view at source ↗
Figure 2
Figure 2. Effectiveness of CN/ACN. (a) shows that our method is effective across various backbones, where 1) non-CID models (e.g., RMLP, iTransformer) exhibit greater improvements from CN and 2) data-independent models (e.g., RMLP, TSMixer), whose parameters do not depend on the input, benefit more from transitioning from CN to ACN. (b) shows forecasting results with and without our methods. tween channels effectively. Furthe… view at source ↗
Figure 3
Figure 3. Overall framework of CN/ACN/PCN. (1) CN employs channel-specific parameters, enabling the model to distinguish between channels. (2) ACN extends CN by adapting its parameters to the input TS by utilizing local parameters, which are attended to with different weights based on the similarity between input channels (i.e., channel similarity). (3) PCN makes CN applicable to multiple datasets with unknown/varying number … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Channel entropy gain by CN. (a) Datasets with higher C show a higher entropy gain. (b) Datasets with higher entropy gain show a higher performance gain (average MSE across four horizons). iTransformer RMLP S-Mamba TSMixer iTransformer RMLP S-Mamba TSMixer Feature Ent. …
Figure 8
Figure 8. Figure 8: Visualization of parameters and channels. (a) shows the t-SNE of the parameters of CN, with four clusters formed. (b) visualizes three channels from each cluster, demonstrating that channels in the same cluster share similar patterns except for those in the 4 th cluste…
Figure 9
Figure 9. Figure 9: Effectiveness of CN/ACN under various L. K = 10 K = 20 K = 50 K = 100 t-SNE *Distortion *Sum of squared distances to cluster centers *MSE 0.460 0.453 0.454 0.454 Redundant prototypes. # Clusters # Clusters # Clusters # Clusters *Avg.MSE of 20 FCST tasks (Baseline: 0.47…
Figure 10
Figure 10. Figure 10: t-SNE & distortion plot of PCN parameters. [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 27 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Ahamed, M. A. and Cheng, Q. Timemachine: A time series is worth 4 mambas for long-term forecasting. In ECAI, 2024

    work page 2024

  3. [3]

    R., and Hinton, G

    Ba, J., Kiros, J. R., and Hinton, G. E. Layer normalization. ArXiv e-prints, pp.\ arXiv--1607, 2016

    work page 2016

  4. [4]

    Mambats: Improved selective state space models for long-term time series forecasting

    Cai, X., Zhu, Y., Wang, X., and Yao, Y. Mambats: Improved selective state space models for long-term time series forecasting. arXiv preprint arXiv:2405.16440, 2024

    arXiv 2024

  5. [5]

    R., Chen, M., Rodrigues, M

    Carson, W. R., Chen, M., Rodrigues, M. R., Calderbank, R., and Carin, L. Communications-inspired projection design with application to compressive sensing. SIAM Journal on Imaging Sciences, 5 0 (4): 0 1185--1212, 2012

    work page 2012

  6. [6]

    Freeway performance measurement system: mining loop detector data

    Chen, C., Petty, K., Skabardonis, A., Varaiya, P., and Jia, Z. Freeway performance measurement system: mining loop detector data. Transportation research record, 1748 0 (1): 0 96--102, 2001

    work page 2001

  7. [7]

    E., Feng, A., Hu, W., Fey, M., Tassiulas, L., Leskovec, J., and Ying, R

    Chen, J., Lenssen, J. E., Feng, A., Hu, W., Fey, M., Tassiulas, L., Leskovec, J., and Ying, R. From similarity to superiority: Channel clustering for time series forecasting. arXiv preprint arXiv:2404.01340, 2024 a

    arXiv 2024

  8. [8]

    O., and Pfister, T

    Chen, S.-A., Li, C.-L., Yoder, N., Arik, S. O., and Pfister, T. Tsmixer: An all-mlp architecture for time series forecasting. TMLR, 2023

    work page 2023

Show all 49 references
  1. [9]

    Learning on bandwidth constrained multi-source data with mimo-inspired dpp map inference

    Chen, X., Li, H., Amin, R., and Razi, A. Learning on bandwidth constrained multi-source data with mimo-inspired dpp map inference. IEEE Transactions on Machine Learning in Communications and Networking, 2024 b

    2024

  2. [10]

    Sequence complementor: Complementing transformers for time series forecasting with learnable sequences

    Chen, X., Qiu, P., Zhu, W., Li, H., Wang, H., Sotiras, A., Wang, Y., and Razi, A. Sequence complementor: Complementing transformers for time series forecasting with learnable sequences. In AAAI, 2025

    2025

  3. [11]

    Injecttst: A transformer method of injecting global information into independent channels for long time series forecasting

    Chi, C., Wang, X., Yang, K., Song, Z., Jin, D., Zhu, L., Deng, C., and Feng, J. Injecttst: A transformer method of injecting global information into independent channels for long time series forecasting. arXiv preprint arXiv:2403.02814, 2024

    2024 arXiv

  4. [12]

    Towards spatio-temporal aware traffic time series forecasting

    Cirstea, R.-G., Yang, B., Guo, C., Kieu, T., and Pan, S. Towards spatio-temporal aware traffic time series forecasting. In 2022 IEEE 38th International Conference on Data Engineering (ICDE), pp.\ 2900--2913. IEEE, 2022

    2022

  5. [13]

    A hybrid residual dilated lstm and exponential smoothing model for midterm electric load forecasting

    Dudek, G., Pe ka, P., and Smyl, S. A hybrid residual dilated lstm and exponential smoothing model for midterm electric load forecasting. IEEE Transactions on Neural Networks and Learning Systems, 33 0 (7): 0 2879--2891, 2021

    2021

  6. [14]

    Units: Building a unified time series model

    Gao, S., Koker, T., Queen, O., Hartvigsen, T., Tsiligkaridis, T., and Zitnik, M. Units: Building a unified time series model. arXiv preprint arXiv:2403.00131, 2024

    2024 arXiv

  7. [15]

    and Dao, T

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

    2023 arXiv

  8. [16]

    The capacity and robustness trade-off: Revisiting the channel independent strategy for multivariate time series forecasting

    Han, L., Ye, H.-J., and Zhan, D.-C. The capacity and robustness trade-off: Revisiting the channel independent strategy for multivariate time series forecasting. arXiv preprint arXiv:2304.05206, 2023

    2023 arXiv

  9. [17]

    B., Ord, J

    Hyndman, R., Koehler, A. B., Ord, J. K., and Snyder, R. D. Forecasting with exponential smoothing: the state space approach. Springer Science & Business Media, 2008

    2008

  10. [18]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Ioffe, S. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015

    2015

  11. [19]

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

    Kim, T., Kim, J., Tae, Y., Park, C., Choi, J.-H., and Choo, J. Reversible instance normalization for accurate time-series forecasting against distribution shift. In ICLR, 2021

    2021

  12. [20]

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

    Lai, G., Chang, W.-C., Yang, Y., and Liu, H. Modeling long-and short-term temporal patterns with deep neural networks. In The 41st international ACM SIGIR conference on research & development in information retrieval, pp.\ 95--104, 2018

    2018

  13. [21]

    Sequential order-robust mamba for time series forecasting

    Lee, Seunghan Hong, J., Park, T., and Lee, K. Sequential order-robust mamba for time series forecasting. arXiv preprint arXiv:2410.23356, 2024

    2024 arXiv

  14. [22]

    Predicting best-selling new products in a major promotion campaign through graph convolutional networks

    Li, C., Jiang, W., Yang, Y., Pan, S., Huang, G., and Guo, L. Predicting best-selling new products in a major promotion campaign through graph convolutional networks. IEEE Transactions on Neural Networks and Learning Systems, 34 0 (11): 0 9102--9115, 2022

    2022

  15. [23]

    Revisiting long-term time series forecasting: An investigation on linear mapping

    Li, Z., Qi, S., Li, Y., and Xu, Z. Revisiting long-term time series forecasting: An investigation on linear mapping. arXiv preprint arXiv:2305.10721, 2023

    2023 arXiv

  16. [24]

    Bi-mamba+: Bidirectional mamba for time series forecasting

    Liang, A., Jiang, X., Sun, Y., and Lu, C. Bi-mamba+: Bidirectional mamba for time series forecasting. arXiv preprint arXiv:2404.15772, 2024

    2024 arXiv

  17. [25]

    Scinet: Time series modeling and forecasting with sample convolution and interaction

    Liu, M., Zeng, A., Chen, M., Xu, Z., Lai, Q., Ma, L., and Xu, Q. Scinet: Time series modeling and forecasting with sample convolution and interaction. In NeurIPS, 2022

    2022

  18. [26]

    itransformer: Inverted transformers are effective for time series forecasting

    Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., and Long, M. itransformer: Inverted transformers are effective for time series forecasting. In ICLR, 2024

    2024

  19. [27]

    Fmamba: Mamba based on fast-attention for multivariate time-series forecasting

    Ma, S., Kang, Y., Bai, P., and Zhao, Y.-B. Fmamba: Mamba based on fast-attention for multivariate time-series forecasting. arXiv preprint arXiv:2407.14814, 2024

    2024 arXiv

  20. [28]

    and Gweon, H

    McLeod, A. and Gweon, H. Optimal deseasonalization for monthly and daily geophysical time series. Journal of Environmental statistics, 4 0 (11): 0 1--11, 2013

    2013

  21. [29]

    Channel-aware low-rank adaptation in time series forecasting

    Nie, T., Mei, Y., Qin, G., Sun, J., and Ma, W. Channel-aware low-rank adaptation in time series forecasting. In CIKM, pp.\ 3959--3963, 2024

    2024

  22. [30]

    H., Sinthong, P., and Kalagnanam, J

    Nie, Y., Nguyen, N. H., Sinthong, P., and Kalagnanam, J. A time series is worth 64 words: Long-term forecasting with transformers. In ICLR, 2023

    2023

  23. [31]

    Solar power data for integration studies

    NREL. Solar power data for integration studies. https://www.nrel.gov/grid/solar-power-data.html, 2006

    2006

  24. [32]

    Deep adaptive input normalization for time series forecasting

    Passalis, N., Tefas, A., Kanniainen, J., Gabbouj, M., and Iosifidis, A. Deep adaptive input normalization for time series forecasting. TNNLS, 31 0 (9): 0 3760--3765, 2019

    2019

  25. [33]

    Pytorch: An imperative style, high-performance deep learning library

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Kopf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., and Chintala, S. Pytorch: An impera...

    2019

  26. [34]

    Certain relations between mutual information and fidelity of statistical estimation

    Prasad, S. Certain relations between mutual information and fidelity of statistical estimation. arXiv preprint arXiv:1010.1508, 2010

    2010 arXiv

  27. [35]

    E., Hinton, G

    Rumelhart, D. E., Hinton, G. E., and Williams, R. J. Learning representations by back-propagating errors. nature, 323 0 (6088): 0 533--536, 1986

    1986

  28. [36]

    and Joy, A

    Thomas, M. and Joy, A. T. Elements of information theory. Wiley-Interscience, 2006

    2006

  29. [37]

    Instance normalization: The missing ingredient for fast stylization

    Ulyanov, D. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016

    2016 arXiv

  30. [38]

    and Hinton, G

    Van der Maaten, L. and Hinton, G. Visualizing data using t-sne. JMLR, 9 0 (11), 2008

    2008

  31. [39]

    N., Kaiser, ., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention is all you need. In NeurIPS, 2017

    2017

  32. [40]

    Is mamba effective for time series forecasting? Neurocomputing, 619: 0 129178, 2025

    Wang, Z., Kong, F., Feng, S., Wang, M., Yang, X., Zhao, H., Wang, D., and Zhang, Y. Is mamba effective for time series forecasting? Neurocomputing, 619: 0 129178, 2025

    2025

  33. [41]

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

    Wu, H., Xu, J., Wang, J., and Long, M. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In NeurIPS, 2021

    2021

  34. [42]

    and He, K

    Wu, Y. and He, K. Group normalization. In ECCV, pp.\ 3--19, 2018

    2018

  35. [43]

    Yu, Y., Chan, K. H. R., You, C., Song, C., and Ma, Y. Learning diverse and discriminative representations via the principle of maximal coding rate reduction. In NeurIPS, 2020

    2020

  36. [44]

    Are transformers effective for time series forecasting? In AAAI, 2023

    Zeng, A., Chen, M., Zhang, L., and Xu, Q. Are transformers effective for time series forecasting? In AAAI, 2023

    2023

  37. [45]

    C-mamba: Channel correlation enhanced state space models for multivariate time series forecasting

    Zeng, C., Liu, Z., Zheng, G., and Kong, L. C-mamba: Channel correlation enhanced state space models for multivariate time series forecasting. arXiv preprint arXiv:2406.05316, 2024

    2024 arXiv

  38. [46]

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

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., and Zhang, W. Informer: Beyond efficient transformer for long sequence time-series forecasting. In AAAI, 2021

    2021

  39. [47]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  40. [48]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  41. [49]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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