acceptodds
Under review as a conference paper at ICLR 2027

CVENet: A Complementary View and Expert Network for Time Series Forecasting with Exogenous Variables

Abstract

Time series forecasting is essential in many domains. Beyond endogenous variables (i.e., target variables), exogenous variables (i.e., covariates) provide additional predictive information over the history, and often over the forecast horizon as well. However, existing methods for time series forecasting with exogenous variables (TSF-X) mainly focus on how endogenous and exogenous variables interact, and fix in advance how the series are processed before and during this interaction, leading to two shortcomings: 1) they handle non-stationarity with a single treatment, either removing the window statistics or keeping them, although the preferable one may shift along the forecast horizon, and 2) they apply the same functional form to every token, as even mixture-of-experts forecasters mostly route tokens among architecturally identical experts. To address these shortcomings, we propose CVENet, a Complementary View and Expert Network for time series forecasting with exogenous variables. Specifically, Complementary View Construction first builds two multi-scale views of the history, with and without the window statistics removed, each carried by its own stack of encoder layers with dual-correlation attention. In every encoder layer, Token-Wise Complementary Experts then route each token within a frequency-domain and a time-domain expert family and mix the two with a token-dependent gate. Finally, Horizon-Wise View Reconciliation combines the two view forecasts with a weight learned for each forecast step. Extensive experiments on twelve real-world datasets and a tokamak dataset demonstrate that CVENet outperforms state-of-the-art methods.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jAWqOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CVENet targets exogenous-aware time series forecasting. The authors argue that prior work has two shortcomings: (i) a single prescribed treatment of non-stationarity is applied before endogenous–exogenous interaction, and (ii) a single functional form (typically an MLP) is applied to every token. To address these, the paper proposes three modules. Complementary View Construction builds an observation view and an instance-normalized view of the history at multiple scales, each carried by its own encoder stack with dual-correlation attention. Horizon-Wise View Reconciliation combines the two view forecasts with a learned per-forecast-step weight. Token-Wise Complementary Experts (TCE) route each token with top-k selection independently within a Fourier/frequency expert family and an MLP/time expert family, mix the two outputs with a token-dependent sigmoid gate, and add an always-on shared expert implemented with a depthwise convolution. Training combines a prediction loss, per-view supervision, correlation-modeling auxiliary losses, and a load-balancing regularizer. Experiments cover twelve public datasets under two exogenous-availability regimes, plus EAST-POINT127, a tokamak electron-density dataset with 40 actuator signals.

Critical strengths
  1. A coherent motivation and unified architecture.

The paper identifies two meaningful challenges in time series forecasting with exogenous variables: choosing how to handle non-stationarity and adapting token transformations to heterogeneous temporal patterns. CVENet addresses both within a unified framework rather than treating them as entirely separate problems. The resulting design is intuitive and potentially useful.

  1. Broad experimental evaluation.

The evaluation covers twelve public datasets, two future-exogenous-variable availability regimes, multiple forecasting horizons, and a tokamak dataset. The inclusion of both conventional forecasting benchmarks and an application involving actuator signals and electron-density measurements is valuable.

  1. Practical relevance.

The framework explicitly addresses the situation in which future exogenous variables are unavailable. This is important in practical forecasting, where covariate forecasts may be unavailable or unreliable at deployment time. The paper therefore targets a meaningful problem rather than a narrowly defined architectural benchmark.

Critical weaknesses
  1. The methodological novelty is not sufficiently distinguished from prior work.

CVENet combines several established ideas: multi-scale trend-seasonal decomposition, instance normalization, dual-correlation attention, mixture-of-experts routing, Fourier-based transformations, auxiliary supervision, and learnable forecast fusion. Although their integration may be useful, the paper does not convincingly establish which aspects constitute fundamental methodological advances rather than a composition of existing techniques.

  1. The experimental design does not fully isolate the sources of improvement.

CVENet introduces substantial architectural complexity. It maintains two encoder stacks, employs multiple expert networks and routers, adds a shared expert, and uses several auxiliary losses. The current ablations do not adequately disentangle the effects of increased parameter count, additional computation, independent view-specific parameters, and the proposed inductive biases.

  1. Statistical reliability of the reported improvements is unclear.

The paper reports competitive average performance, including reductions of 3.9% in MSE and 3.0% in MAE relative to DAG in the setting with future exogenous variables. These results are promising, but the aggregate improvements are not large enough to dismiss the possibility of run-to-run variation without further evidence.

  1. The proposed mechanisms are not sufficiently validated.

The paper motivates frequency-domain experts through their periodic inductive bias and horizon-specific reconciliation through the changing relevance of observations and standardized views. However, the reported results primarily establish predictive performance, rather than directly validating these explanations.

  1. Evaluation under missing future exogenous variables raises questions about error propagation.

When future exogenous variables are unavailable, CVENet uses predictions from its temporal branch as inputs to the channel branch. This is a reasonable design, but it introduces a potential error-propagation mechanism: errors in the predicted covariates may affect the final endogenous forecasts.

  1. Efficiency and reproducibility require stronger evidence.

The manuscript reports training on a single NVIDIA RTX 4090 and states that source code will be released after acceptance. However, the paper does not sufficiently establish the relative training cost, inference latency, memory consumption, or parameter efficiency of CVENet compared with strong baselines. These measurements are important because the proposed architecture duplicates view-specific processing and introduces multiple expert families.

Questions for authors

Refer to weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer jAWq

Dear Reviewer jAWq,

Thank you for your detailed review. Below we address all six points (additions marked in blue in the PDF).

W1: TSF-X Formulation vs. Existing Primitives

As noted in General Response Point 1 (Sections 1, 2.2, 2.3, 3.2), TSF-X couples two choices treated separately in prior work: keeping observation and standardized views in separate stacks (unlike single-normalization Autoformer, FEDformer, TimeMixer, FreTS, and FITS) and mixing parallel Fourier and MLP families per post-attention token (unlike homogeneous MLPs in DUET or covariate-free single-family layers in MoFE-Time). Across all 48 public benchmarks via Algorithm 1 (Appendix B.6, Table 10), FreTS (0.454), TimeMixer (0.476), FEDformer (0.563), and Autoformer (0.630) incur 42.1%–97.3% higher MSE than CVENet (0.319, a 29.6%–49.3% reduction) with future covariates, where CVENet ranks 1st across all 15 models; without future covariates (0.467 vs. 0.533–0.868), CVENet beats all four in dataset-averaged MSE and MAE on all 12 datasets.

W2: Isolating Capacity, Encoders, Experts, and Losses

Across all 24 public benchmarks (Section 4.3, Figure 4; Appendix B.5, Table 9), we tested matched-scale homogeneous controls and widened dd, FFN width, or depth to match CVENet's size:

VariantParamsMSEvs. CVENetWins
CVENet1.00×\mathbf{1.00\times}0.319----
DAG (widened)1.02×1.02\times0.383+19.9%23/24
Single view (widened)1.03×1.03\times0.417+30.5%24/24
Standard FFN (widened)1.02×1.02\times0.369+15.7%23/24
Two observation views1.00×1.00\times0.356+11.7%23/24
Two standardized views1.00×1.00\times0.347+8.7%19/24
Fourier-only (E,kE, k fixed)0.95×0.95\times0.363+13.8%23/24
MLP-only (E,kE, k fixed)1.05×1.05\times0.345+8.0%20/24
w/o Shared expert0.96×0.96\times0.349+9.2%24/24

Sharing one encoder stack across both views raises MSE by 10.2% (24/24 benchmarks), widening DAG raises its own MSE from 0.332 to 0.383 (+15.2%, losing 22/24), and setting β=0\beta = 0, γ=0\gamma = 0, or λ=0\lambda = 0 raises MSE to 0.368 (+15.3%, 22/24), 0.355 (+11.1%, 23/24), and 0.339 (+6.2%, 16/24).

W3: Statistical Reliability Across Five Seeds

Over five paired seeds (2021–2025) on identical hardware and loaders across all 24 benchmarks, CVENet beats DAG in mean MSE on 19/24 benchmarks (sign test p=0.0033p = \mathbf{0.0033}, winning 79/120 paired runs) and in mean MAE on 16/24 with 2 ties (p=0.026p = \mathbf{0.026}), alongside 20/24 wins in Table 7 (p=0.0008p = 0.0008), with tight dispersion (e.g., PJM: 0.055±0.001 vs. 0.057±0.001 at F=24F = 24 and 0.117±0.009 vs. 0.130±0.005 at F=360F = 360; Sdwpfm1: 0.326±0.013 vs. 0.363±0.010 and 0.496±0.019 vs. 0.523±0.017; Colbún: 0.056±0.004 vs. 0.061±0.001 and 0.144±0.013 vs. 0.159±0.035). On 51-variable EAST-POINT127 (Table 2), CVENet wins 19/20 entries, cutting MSE by 14.3% over DAG and 27.7%–59.7% over TimeXer, TiDE, and TFT.

W4: Direct Validation of g(x)g(x) and ωh\omega_h

  • Token gate g(x)g(x) (Sections 3.4, 4.3): Fixing g=0.5g = 0.5 in training and inference (+6.2%, 18/24), replacing wg⊤xw_g^\top x with a learned layer scalar (+8.5%, 20/24), or forcing g=0g = 0 (+8.9%) or g=1g = 1 (+55.7%) at inference raises 24-benchmark MSE; on a 3-seed synthetic switching test, learned gating (0.281±0.004) beats time-only (+46.2%) and frequency-only (+40.4%).
  • Horizon weights ωh\omega_h (Section 4.3, Figure 3): Learned weights track the lower-error view across steps—high on Sdwpfm1 (0.62–0.72, ρ=+0.66\rho = +0.66), rising on NP (0.42→0.530.42 \to 0.53, ρ=+0.75\rho = +0.75), and falling on FR (0.51→0.370.51 \to 0.37, ρ=+0.67\rho = +0.67)—beating both views simultaneously on 20/24 benchmarks (-9.3% and -7.7% mean MSE).

W5 & W6: Error Propagation, Covariate Noise, and Efficiency

  • Missing and noisy covariates: Under fixed history-only weights across all 24 benchmarks (Section 4.3), feeding predicted Y^exo\hat{Y}^{\text{exo}} into the channel branch lowers MSE by 6.3% (0.498 →\to 0.467, 21/24 wins) vs. disabling it (α=1\alpha = 1). Under test-time N(0,s2)\mathcal{N}(0, s^2) noise on YexoY^{\text{exo}} (s=0.1,0.3,0.5s = 0.1, 0.3, 0.5; Appendix B.7, Figure 13, Table 11), CVENet's MSE rise (4.0%, 6.5%, 10.4%) is below half DAG's (8.7%, 15.8%, 20.7%).
  • Efficiency: On one RTX 4090 (B=64B = 64) across NP (F=24,360F = 24, 360) and Energy (F=24F = 24), CVENet runs in 0.51–0.63 ms/win inference and 0.10–0.13 s/step training with 62–349 MiB peak memory (4.1×4.1\times–5.7×5.7\times fewer FLOPs/win than GCGNet at F=24F = 24; 11.1×11.1\times–23.7×23.7\times fewer than TimeMixer), and 4.71 ms/win with 7.3 GiB peak memory on EAST-POINT127 (96→72096 \to 720).

We are happy to discuss any of these results further.

I thank the authors for the detailed response and for the substantial additional experiments. I acknowledge that most of my empirical concerns have been addressed or substantially mitigated.

However, my primary concern remains the methodological novelty (W1), and the response does not fully resolve it. The response clarifies the composition and demonstrates its effectiveness, but does not establish a fundamental methodological advance beyond a well-motivated arrangement of existing techniques. The empirical evidence supports effectiveness, not novelty.

To reconsider my score, I would need some theoretical support for the methodological design, even under simplifying assumptions (e.g., linear models or controlled synthetic settings). Specifically, a focused analysis showing why the two complementary views together strictly expand the representable function class, or reduce the achievable error, relative to either view alone would directly address my concern.

AuthorsOpenReview ↗

Follow-up Response to Reviewer jAWq

Dear Reviewer jAWq,

Thank you for your follow-up and for noting that most of your empirical concerns are resolved. We agree on W1: empirical gains show effectiveness, while novelty requires proving why the coupled design is necessary.

1. Structural Dilemma in TSF-X

Across power, weather, hydro, and plasma control (Section 4.1; Appendix A.1), single designs hit three limits:

  1. Level relaxation vs. σ\sigma-scaled gain. Over step hh, memory of μ\mu fades while covariate gain scales with σ\sigma (grid clearing, hydro storage, or plasma transport lag; Wan et al., 2026). Observation views re-weight μ\mu under fixed slope; standardization (Kim et al., 2022) captures σ\sigma-scaled gain yet locks 1⋅μ1\cdot\mu, adding drift.
  2. Coupled views vs. naive stacking. Duplicating a view stays in one subspace, while unanchored fusion (γ=0\gamma=0) leaves branches unidentifiable, requiring view loss and step-wise weights.
  3. Cycles plus jumps. Attention mixes cycles and covariate steps in one token stream (Figures 5–6), requiring token-wise Fourier and MLP routing.

2. Function-Class and Achievable-Risk Analysis

For x∈RTx \in \mathbb{R}^T (T≥2T \ge 2) with mean μ\mu, std σ>0\sigma > 0 (σˉ=E[σ]\bar\sigma=E[\sigma]), z=(x−μ1)/σz=(x-\mu\mathbf{1})/\sigma, and uh∈Ru_h \in \mathbb{R}, linear predictors (Sections 3.3, 3.5) form:

Fo={wh⊤x+βhuh+ch},Fs={μ+σ(w~h⊤z+β~huh+c~h)}.F_o = \lbrace w_h^\top x + \beta_h u_h + c_h\rbrace, \quad F_s = \lbrace\mu + \sigma(\tilde w_h^\top z + \tilde\beta_h u_h + \tilde c_h)\rbrace.

Proposition 1 (Function-Class Expansion). Over zero-mean shapes V0={v⊤(x−μ1)}V_0=\lbrace v^\top(x-\mu\mathbf{1})\rbrace (dim⁡T−1\dim T-1) for σ>0\sigma > 0, FoF_o (dim⁡T+2\dim T+2) adds μ,uh,1\mu, u_h, 1 with ∂f/∂uh=βh\partial f/\partial u_h = \beta_h (ρhμ∈Fo∖Fs\rho^h\mu \in F_o \setminus F_s, ρh≠1\rho^h \neq 1), while FsF_s (affine dim⁡T+1\dim T+1) locks f(x+δ1,uh)=f(x,uh)+δf(x+\delta\mathbf{1}, u_h)=f(x, u_h)+\delta and adds σuh,σ\sigma u_h, \sigma (μ+σuh∈Fs∖Fo\mu+\sigma u_h \in F_s \setminus F_o). Homogeneous stacks collapse (ωhFv+(1−ωh)Fv=Fv\omega_h F_v + (1-\omega_h)F_v = F_v), while for ωh∈(0,1)\omega_h \in (0,1) reconciling adds μ,uh,1,σuh,σ\mu, u_h, 1, \sigma u_h, \sigma (dim⁡T+4\dim T+4):

Fo∪Fs⊊ωhFo+(1−ωh)Fs.F_o \cup F_s \subsetneq \omega_h F_o + (1-\omega_h)F_s.

Proposition 2 (Risk Reduction and ωh∗\omega_h^\ast). Let yh=ρhμ+θh⊤(x−μ1)+κhσuh+ξhy_h=\rho^h\mu+\theta_h^\top(x-\mu\mathbf{1})+\kappa_h\sigma u_h+\xi_h (ρ∈(0,1)\rho \in (0,1), 1⊤θh=0\mathbf{1}^\top\theta_h=0, κh≠0\kappa_h \neq 0; independent zz, σ>0\sigma > 0 of variance νσ2>0\nu_\sigma^2 > 0, zero-mean μ,uh,ξh\mu, u_h, \xi_h of variances νμ2>0,1,σξ2\nu_\mu^2 > 0, 1, \sigma_\xi^2), ah=(1−ρh)2νμ2a_h=(1-\rho^h)^2\nu_\mu^2, and bh=κh2νσ2b_h=\kappa_h^2\nu_\sigma^2. Single-view MSE floors are Ro∗(h)=bh+σξ2R_o^\ast(h)=b_h+\sigma_\xi^2 on FoF_o and Rs∗(h)=ah+σξ2R_s^\ast(h)=a_h+\sigma_\xi^2 on FsF_s (and nonlinear μ+σΨh(z,uh)\mu+\sigma\Psi_h(z, u_h), as zz strips μ\mu). E[yh∣x,uh]∈ωhFo+(1−ωh)FsE[y_h \mid x, u_h] \in \omega_h F_o + (1-\omega_h)F_s hits σξ2\sigma_\xi^2, yet γ=0\gamma=0 leaves (fo,fs,ωh)(f_o, f_s, \omega_h) unidentifiable; under Lview\mathcal L_{\text{view}} (Eqs. 11, 14), optimal y^ho,∗,y^hs,∗\hat y_h^{o,\ast}, \hat y_h^{s,\ast} have orthogonal errors κh(σ−σˉ)uh\kappa_h(\sigma-\bar\sigma)u_h and (ρh−1)μ(\rho^h-1)\mu, so y^h=ωhy^ho,∗+(1−ωh)y^hs,∗\hat y_h=\omega_h\hat y_h^{o,\ast}+(1-\omega_h)\hat y_h^{s,\ast} (Eq. 10) uniquely gives:

ωh∗=ahah+bh∈(0,1),R∗(h)=ahbhah+bh+σξ2<min⁡{Ro∗(h),Rs∗(h)},\omega_h^\ast = \frac{a_h}{a_h + b_h} \in (0,1), \quad R^\ast(h) = \frac{a_h b_h}{a_h + b_h} + \sigma_\xi^2 < \min\lbrace R_o^\ast(h), R_s^\ast(h)\rbrace,

and shared ωˉ\bar\omega (Table 3c) adds 1F∑h=1F(ah+bh)(ωˉ−ωh∗)2>0\frac{1}{F}\sum_{h=1}^F (a_h+b_h)(\bar\omega-\omega_h^\ast)^2 > 0 when ωh∗\omega_h^\ast varies with hh.

Dual-family experts (Eqs. 5–9). An MM-term Fourier sum fits MM tones yet has Ω(1/M)\Omega(1/M) Gibbs MSE on jumps; a GELU MLP fits steps/ramps in O(1)O(1) units without periodic bias (Dong et al., 2025), motivating gate gg (Eq. 8).

3. Synthetic Check and Empirical Alignment

Over 5 seeds (2021–2025; T=96T=96, F=24F=24, ρ=0.9\rho=0.9, νμ=1\nu_\mu=1, σ∼U(0.5,2.5)\sigma \sim U(0.5,2.5), κh=1\kappa_h=1, σξ2=0.05\sigma_\xi^2=0.05; 50k/50k), OLS confirms both:

VariantTheory5-Seed MSE
Observation view (FoF_o)0.3830.383 ± 0.001
Standardized view (FsF_s)0.5360.538 ± 0.001
Two observation views0.3830.384 ± 0.001
Two standardized views0.5360.539 ± 0.001
Shared ωˉ\bar\omega0.2480.248 ± 0.001
Step-wise ω^h\hat\omega_h0.2260.227 ± 0.001 (-40.9%)

Learned ω^h\hat\omega_h rises from 0.029 (h=1h=1) to 0.716 (h=24h=24), matching ωh∗\omega_h^\ast (0.029→0.7180.029 \to 0.718, r=0.9999r=0.9999). This ωh∗\omega_h^\ast ratio explains Figure 3—NP (0.42→0.530.42 \to 0.53, rising aha_h), Sdwpfm1 (0.620.62–0.720.72, ah>bha_h > b_h), FR (0.51→0.370.51 \to 0.37, rising bhb_h)—plus homogeneous (+8.7%–11.7%), widened single-view (+30.5%; Figure 4, Table 9), γ=0\gamma=0 (+11.1%; Figure 11a, Table 9), and single-family switching (+40.4%–46.2%; Section 4.3) losses, and EAST-POINT127 horizon gains (96 ms: 0.024/0.026 →\to 720 ms: 0.289/0.331; Table 2).

We will add a summary in Section 3.5 and full proofs in Appendix C. Thank you again for your feedback.

Reviewer_tidzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CVENet, a deep learning architecture for time series forecasting with exogenous variables. The authors identify two main limitations in existing models: 1) the rigid handling of non-stationarity using either fully normalized or unnormalized data, ignoring the possibility that statistical relevance changes across the forecasting horizon; and 2) the use of homogeneous functional forms across all tokens in Transformer/MoE architectures. The core design of CVENet:

  1. Decomposing series into an Observation View and a Standardized View at multiple scales, and reconciling their forecasts step-by-step using learnable horizon weights.

  2. A dual-family MoE that routes tokens into frequency-domain (Fourier analysis) and time-domain (MLP) expert families with token-conditioned mixing.

The model is evaluated on 12 public benchmarks and a tokamak plasma dataset (EAST-POINT127) under both exogenous-available and exogenous-unavailable regimes, demonstrating competitive empirical metrics.

Critical strengths

1.The paper evaluates the model on 12 benchmark datasets across multiple domains (electricity, hydrology, wind power) as well as a large-scale real-world scientific dataset (EAST-POINT127 tokamak fusion). The reported results consistently outperform existing baselines such as DAG and TimeXer.

2.The observation that historical window statistics are more reliable for near-term forecasting steps than long-term future horizons is conceptually sound and provides an insight perspective on multi-step reconciliation.

Critical weaknesses

1.For multi-scale decomposition via moving average kernels has been widely used in prior literature (e.g., Autoformer, FEDformer), and multi-scale deconstruct/mix mechanisms have been extensively established in architectures like TimeMixer. For token-wise expert , fourier analysis layers and standard MLPs within an MoE gating framework. Combining time-domain and frequency-domain representations via soft gating is a well-trodden direction in time series forecasting (e.g., FreTS, FITS, MoFE-Time). The overall framework resembles a combination of existing architectural primitives rather than a groundbreaking methodological innovation.

2.The paper claims that Token-Wise Complementary Experts dynamically allocate periodic tokens to frequency experts and non-periodic tokens to time-domain experts (Figure 1b, Section 3.4). However, there is no direct empirical or theoretical proof supporting this claim.

Questions for authors

1.Does the gate gg truly respond to frequency-domain periodicity, or is it merely acting as an unconstrained parameter regularizer?

2.How does CVENet compare conceptually and empirically with TimeMixer (which also operates on multi-scale past/future mixing) and recent time-frequency MoE models (such as MoFE-Time)?

3.Could authors provide a detailed comparison of FLOPs, model parameter count, and inference latency between CVENet, DAG, and TimeMixer?

4 comments
AuthorsOpenReview ↗

Response to Reviewer tidz

Dear Reviewer tidz,

Thank you for your questions on multi-scale and time-frequency baselines, direct tests of g(x)g(x), and computational cost. Below we address all three points, with additions marked in blue in the revised PDF.

W1 & Q2: Comparison with TimeMixer, FreTS, and MoFE-Time

Sections 1, 2.2, 2.3, and 3.2 contrast CVENet with these two lines of work:

  1. Multi-scale and frequency models (Autoformer, FEDformer, TimeMixer, FreTS, FITS): Decomposition and frequency filtering are well established, but these models (including FITS, a ∼104\sim 10^4-parameter channel-independent low-pass filter) use one upfront normalization without native future-covariate interaction. Under external drivers, how long historical levels stay informative shifts across the horizon: sharing one encoder stack across both views raises 24-benchmark MSE by 10.2% across all 24/24 benchmarks, whereas reconciling separate stacks via ωh\omega_h beats both single views simultaneously on 20/24 benchmarks, lowering mean MSE by 9.3% and 7.7% (Section 4.3, Figure 3).
  2. Time-frequency MoE (MoFE-Time): Cross-variable attention mixes periodic cycles with abrupt covariate ramps in one token stream. Unlike MoFE-Time, which pretrains on covariate-free univariate series with one fixed expert type per layer, CVENet routes each post-attention token through parallel Fourier and MLP families. Restricting routed experts (E,kE, k fixed) to Fourier layers (+13.8% MSE, losing 23/24) or MLPs (+8.0%, losing 20/24), or widening DAG to match CVENet's size (+19.9% MSE), degrades accuracy (Section 4.3, Figure 4; Appendix B.5, Table 9).

Across all 48 public benchmarks under identical inputs (via Algorithm 1 when YexoY^{\text{exo}} is given; Appendix B.6, Table 10), CVENet outperforms all four added baselines:

Modelw/ Future MSEw/ Future MAEw/o Future MSEw/o Future MAE
CVENet0.3190.3300.4670.406
FreTS0.454 (+42.1%)0.4480.533 (+14.2%)0.481
TimeMixer0.476 (+49.1%)0.4580.554 (+18.8%)0.464
FEDformer0.563 (+76.5%)0.5350.768 (+64.6%)0.571
Autoformer0.630 (+97.3%)0.5620.868 (+86.1%)0.639

Without future covariates, CVENet beats all four in dataset-averaged MSE and MAE across all 12 datasets; with future covariates, it ranks 1st overall across all 15 compared models (mean MSE/MAE ranks 2.29 / 2.17 vs. DAG's 3.13 / 2.96).

W2 & Q1: Does g(x)g(x) Act Only as an Unconstrained Parameter Regularizer?

We clarified in Section 3.4 (Equation 8) that g(x)=Sigmoid(wg⊤x)g(x) = \mathrm{Sigmoid}(w_g^\top x) gates post-attention latent tokens rather than raw series, and ran two direct validation tests (summarized in Section 4.3):

  • 24-benchmark interventions: Across all 24 public benchmarks, fixing g=0.5g = 0.5 throughout training and inference raises average MSE by 6.2% (degrading 18/24 benchmarks), replacing wg⊤xw_g^\top x with a learned layer-wise scalar raises MSE by 8.5% (degrading 20/24), and forcing time-only (g=0g = 0) or frequency-only (g=1g = 1) routing at inference raises MSE by 8.9% or 55.7%, confirming that gg must be both learned and conditioned on xx.
  • Synthetic switching test (3 seeds): On a 40,000-step series alternating multi-harmonic periodic and non-periodic drift/AR regimes (168→24168 \to 24, seeds 2021–2023), learned gating reaches 0.281 ± 0.004 MSE, beating forced time-only (0.410 ± 0.023, +46.2%) and frequency-only (0.394 ± 0.014, +40.4%).

Q3: FLOPs, Parameters, and Latency vs. DAG and TimeMixer

We profiled cost across variable counts and horizon lengths on one RTX 4090 (B=64B = 64 for CVENet/DAG, B=32B = 32 for TimeMixer):

SettingModelParams (M)MFLOPs/winInfer (ms/win)MSE
NP (24, 3 vars)CVENet1.448.500.510.181
NP (24, 3 vars)DAG0.492.290.120.202
NP (24, 3 vars)TimeMixer0.21129.780.190.256
Energy (24, 6 vars)CVENet1.8023.360.580.070
Energy (24, 6 vars)DAG0.565.360.160.079
Energy (24, 6 vars)TimeMixer0.21259.500.270.139
NP (360, 3 vars)CVENet16.7368.820.630.503
NP (360, 3 vars)DAG6.0617.870.120.521
NP (360, 3 vars)TimeMixer3.591632.700.270.682

Across all three settings, CVENet achieves the best MSE while running in 0.51–0.63 ms/win inference with 62–349 MiB peak GPU memory, using fewer parameters and 4.1×\mathbf{4.1\times}–5.7×\mathbf{5.7\times} fewer FLOPs/win than GCGNet at F=24F = 24 (1.44M / 1.80M vs. 2.34M / 2.51M) and 11.1×\mathbf{11.1\times}–23.7×\mathbf{23.7\times} fewer FLOPs/win than TimeMixer.

We are happy to clarify any part of these results during discussion.

Dear Reviewer tidz,

As the author-reviewer discussion deadline is approaching, we wanted to kindly check whether our response and the revised manuscript have addressed your questions and concerns. Please let us know if anything remains unclear or if any further clarification would be helpful—we would be glad to discuss further before the window closes.

Thank you again for your time and helpful feedback.

Best regards,
The Authors

Thank you for your response. My concerns have been addressed. I decide to raise my score.

Dear Reviewer tidz,

Thank you very much for reviewing our response and for raising your score. We truly appreciate your thoughtful comments and suggestions, which have helped us make both the empirical comparisons and the presentation much clearer.

Best regards,
The Authors

Reviewer_kbWGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CVENet addresses time-series forecasting with historical and, when available, future exogenous variables. It constructs two multi-scale views that retain or remove-and-restore window statistics, processes them with separate dual-correlation encoders, and uses token-dependent mixtures of Fourier-analysis and MLP expert families with a shared expert. A learned weight for each forecast step combines the two view forecasts; these horizon weights are shared across inputs. Experiments cover twelve public datasets under two exogenous-input regimes and the EAST-POINT127 tokamak dataset, together with component ablations and sensitivity analyses.

Critical strengths
  1. The design connects two concrete modeling choices to testable forecasting behavior. The paper makes both the statistical treatment of a window and the transformation applied to a token explicit modeling choices. The horizon-wise reconciliation is particularly well isolated: Table 3 compares it with one learned weight shared across the horizon, changing only F−1 scalar parameters. The consistent improvement in this comparison provides direct evidence for step-dependent fusion, beyond an increase in backbone capacity.

  2. The evaluation includes a substantive scientific application and different information regimes. The public-data results are complemented by predicting eleven electron-density channels from forty actuator signals on EAST. Appendix A.1 separates training, validation, and test data at discharge boundaries, avoiding shared discharges across subsets. Table 2 reports a lower average MSE than DAG, and Table 4 evaluates the history-only exogenous setting as well. Together, these results support practical value across more than one narrowly configured benchmark.

Critical weaknesses
  1. The evidence does not fully distinguish complementary mechanisms from additional capacity. Removing the standardized view also removes an independently parameterized encoder; Appendix B.5 itself notes that the two views share no parameters, so parameters and computation grow with the number of views. Replacing TCE with an FFN likewise changes both expert capacity and routing regularization. These ablations therefore do not isolate the advantage of complementary views or heterogeneous expert functions over comparably sized alternatives. This matters because the public-data average improvement over DAG is approximately 3.9% in MSE and 3.0% in MAE, without seed-level uncertainty in the main tables. Parameter counts, runtime, and a representative matched-capacity control would make the architectural claim more convincing.

  2. The deployment meaning of future exogenous inputs requires more precise qualification. Section 4.1 states that future exogenous variables in all public datasets are either known in advance or can be forecast with high accuracy. Appendix A.1, however, describes forecast-derived EPF features, ERA5 reanalysis weather for the wind-power datasets, precipitation and tributary inflow for Colbún and Rapel without stating whether these are observations or forecasts, and actuator signals for EAST. These are different information sources: an archived forecast, a reanalysis or observed value, and a control signal are not automatically available under the same operational conditions. Giving every method the same inputs can make the offline comparison fair while still leaving its deployment interpretation conditional. This limits how broadly the reported advantages can be understood as improvements achievable at forecast time.

Questions for authors
  1. Targeted validation of complementarity: Is there a comparison with two identically treated views or homogeneous experts at similar parameter count and training cost? Please provide any existing representative result, along with parameter/runtime and seed-variability information, that separates complementary design from extra capacity.

  2. Clarification of information availability: For each exogenous-data category, are future inputs archived forecasts available at the forecast origin, precommitted control plans, or subsequently observed/reanalyzed values? For EAST, please also confirm that sampled windows do not cross discharge boundaries within a split. Which conclusions are intended to hold under operationally available inputs, and which under idealized future-input access?

5 comments
AuthorsOpenReview ↗

Response to Reviewer kbWG

Dear Reviewer kbWG,

We appreciate your close reading of our parameter isolation in Table 3 and of the discharge-boundary split on EAST-POINT127. Below we address both of your questions, with corresponding additions marked in blue in the revised manuscript.

W1 & Q1: Separating Complementarity from Capacity, Seed Variability, and Cost

Given the 3.9% MSE / 3.0% MAE mean margin over DAG on the 2–6-covariate public series noted in our General Response (up to 12.9% MSE on PJM and 16.9% MAE on Colbún, alongside 14.3% MSE on EAST-POINT127), we ran matched-scale identical-view and homogeneous-expert controls (0.95×0.95\times–1.05×1.05\times CVENet's parameters; Section 4.3, Figure 4) together with widened equal-parameter baselines (Appendix B.5, Table 9) across all 24 public benchmarks:

VariantParamsMean MSEvs. CVENetWins
CVENet1.00×\mathbf{1.00\times}0.319----
DAG (widened)1.02×1.02\times0.383+19.9%23/24
Single view (widened)1.03×1.03\times0.417+30.5%24/24
Standard FFN (widened)1.02×1.02\times0.369+15.7%23/24
Observation + Observation1.00×1.00\times0.356+11.7%23/24
Standardized + Standardized1.00×1.00\times0.347+8.7%19/24
Fourier experts only (E,kE, k fixed)0.95×0.95\times0.363+13.8%23/24
MLP experts only (E,kE, k fixed)1.05×1.05\times0.345+8.0%20/24

Scaling width alone hurts generalization: widening DAG to match CVENet's parameter count raises DAG's own MSE from 0.332 to 0.383 (+15.2%, losing 22/24 benchmarks to original DAG), whereas reconciling CVENet's two complementary views outperforms both individual views simultaneously on 20/24 benchmarks (Section 4.3, Figure 3).

To verify statistical reliability, we ran five paired random seeds (2021–2025) on identical hardware and data loaders across all 24 public benchmarks: CVENet outperforms DAG in mean MSE on 19 of 24 benchmarks (one-sided sign test p=0.0033p = \mathbf{0.0033}, winning 79 of 120 paired runs) and in mean MAE on 16 of 24 with 2 ties (p=0.026p = \mathbf{0.026}), alongside 20 of 24 wins in Table 7 (p=0.0008p = 0.0008), with tight dispersion across domains (e.g., PJM: 0.055 ± 0.001 vs. 0.057 ± 0.001 at F=24F = 24 and 0.117 ± 0.009 vs. 0.130 ± 0.005 at F=360F = 360; Sdwpfm1: 0.326 ± 0.013 vs. 0.363 ± 0.010 and 0.496 ± 0.019 vs. 0.523 ± 0.017; Colbún: 0.056 ± 0.004 vs. 0.061 ± 0.001 and 0.144 ± 0.013 vs. 0.159 ± 0.035).

Across NP (F=24,360F = 24, 360) and Energy (F=24F = 24) on one RTX 4090 (B=64B = 64), CVENet runs in 0.51–0.63 ms/win inference and 0.10–0.13 s/step training with 62–349 MiB peak GPU memory, using 4.1×\mathbf{4.1\times}–5.7×\mathbf{5.7\times} fewer FLOPs/win than GCGNet at F=24F = 24 and 11.1×\mathbf{11.1\times}–23.7×\mathbf{23.7\times} fewer FLOPs/win than TimeMixer. On 51-variable EAST-POINT127 (96→72096 \to 720), inference takes 4.71 ms/win with 7.3 GiB peak memory against a 720 ms physical horizon.

W2 & Q2: Exogenous Availability, Operational Regimes, and EAST Shot Boundaries

We revised Section 4.1 and Appendix A.1 to clarify operational vs. idealized future-input regimes across all 13 datasets:

  1. Archived day-ahead forecasts (NP, PJM, BE, FR, DE): Load and wind/solar forecasts issued before market clearing; Table 1 holds directly for operational day-ahead deployment.
  2. Measured or reanalysis covariates (Sdwpf ERA5 weather, Colbún/Rapel precipitation and inflow, Energy non-thermal generation): Following GCGNet and DAG, Table 1 tests conditional response under idealized physical drivers. Operational deployment without clean future observations is evaluated in Table 4 (history-only, where CVENet ranks 1st on 7/12 datasets) and in Appendix B.7 (Figure 13 and Table 11), where adding N(0,s2)\mathcal{N}(0, s^2) noise to standardized YexoY^{\text{exo}} at s=0.1,0.3,0.5s = 0.1, 0.3, 0.5 raises CVENet's MSE by only 4.0%, 6.5%, and 10.4% vs. DAG's 8.7%, 15.8%, and 20.7%.
  3. Tokamak actuator trajectories (EAST-POINT127): Constructed and validated in collaboration with plasma-physics and control specialists on the EAST tokamak, the 40 PCS channels record actuator trajectories predicting the 11-channel electron-density response for pulse design and MPC rollout evaluation. Train (67 discharges), validation (16), and test (44) splits divide strictly at discharge boundaries, and restricting evaluation exclusively to strictly intra-discharge windows yields virtually identical accuracy (e.g., 0.0246 / 0.0779 vs. 0.0245 / 0.0780 MSE/MAE at F=96F = 96).

Thank you again for these two questions, which improved both our experimental controls and our dataset documentation.

Dear Reviewer kbWG,

As the author-reviewer discussion deadline is approaching, we wanted to kindly check whether our response and the revised manuscript have addressed your questions and concerns. Please let us know if anything remains unclear or if any further clarification would be helpful—we would be glad to discuss further before the window closes.

Thank you again for your time and helpful feedback.

Best regards,
The Authors

Thank you for the detailed response and the revised manuscript.

The matched-scale controls using identical views or homogeneous experts provide useful evidence for complementarity. I find these more informative than the widened DAG baseline, whose performance worsens after widening. Building on your response to my original concerns, I would appreciate a few further clarifications.

The five-seed comparison and cost measurements in your response are helpful. Could you provide the complete results, including seed variability, alongside parameter counts, FLOPs, and latency? The reported NP results at a 24-step horizon indicate roughly 3× the parameters and 4× the inference time of DAG, so this trade-off should be considered alongside the accuracy gains.

Your clarification of archived forecasts versus measured or reanalysis inputs is also helpful. Averaging the reported dataset MSEs, I obtain a reduction over DAG of about 1.1% on the five EPF datasets with archived forecasts, compared with 5.9% on the other seven datasets. Could you report these groups separately and clarify which conclusions apply under operationally available future inputs?

For EAST, the reported intra-discharge evaluation is reassuring. To clarify the windowing point in my original review, can training or test windows cross discharge boundaries within a split? If so, what fraction does this affect? Does the reported comparison only restrict test windows, or were the models also retrained using intra-discharge windows only? Please clarify the sampling rule and report the corresponding results.

These clarifications will inform my further evaluation of the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

Thank you for your time and constructive feedback. We appreciate your positive comments on our evaluation across 12 public benchmarks and the 51-variable EAST-POINT127 tokamak dataset, as well as the parameter isolation of step-wise view reconciliation in Table 3. We have run new controlled experiments and uploaded a revised PDF with all additions marked in blue.

Before our individual replies, we clarify two themes shared across the reviews:

1. TSF-X formulation vs. generic decomposition and MoE blocks (Reviewers tidz and jAWq; Sections 1, 2.2, 2.3, and 3.2). Exogenous-aware forecasting poses two coupled difficulties absent in univariate decomposition or standard MoE layers. First, external drivers alter how long the input window statistics stay a reliable reference across the horizon: removing and restoring (μ,σ)(\mu, \sigma) anchors predictions to the current window level, whereas passing observed levels through learnable trend maps lets the model re-express how the level evolves. Keeping observation and standardized multi-scale views in separate dual-correlation encoder stacks and reconciling them per step via ωh\omega_h avoids committing all horizons to one upfront treatment. Second, cross-variable attention places periodic cycles alongside abrupt exogenous ramps or actuator steps in one token sequence, motivating parallel Fourier and perceptron expert families mixed per post-attention token via g(x)g(x).

2. Improvement scale on public benchmarks vs. EAST-POINT127 (Reviewers kbWG and jAWq; Section 4.2). We retain the 12 public TSF-X datasets and splits from EPF, TimeXer, GCGNet, and DAG unchanged. Because each series has only 2 to 6 covariates and recent TSF-X baselines already cluster closely on day-ahead market and weather series, aggregate public-benchmark gains move in steady ~3–4% steps: DAG (ICML 2026) reduced average MSE by 3.4% and MAE by 2.6% over GCGNet (ICLR 2026), and CVENet lowers MSE by another 3.9% and MAE by 3.0% over DAG (7.2% cumulative MSE gain over GCGNet). On 51-variable EAST-POINT127 (40 control actuators driving 11 electron-density channels, developed and validated with plasma-physics and control specialists on EAST), coupled multi-channel dynamics leave much wider headroom: CVENet ranks 1st in 19 of 20 entries, cutting MSE by 14.3% over DAG and 27.7%–59.7% over TimeXer, TiDE, and TFT (Table 2).

Index of Manuscript Revisions

  • Section 4.1 & Appendix A.1: Exogenous variable sources across all 13 datasets and actuator-conditioned density forecasting setup on EAST-POINT127.
  • Section 3.4 & Section 4.3 (Figure 3): Contextualized token gating g(x)g(x) and step-wise view reconciliation (ωh\omega_h) profiles across forecast horizons.
  • Section 4.3 (Figure 4) & Appendix B.5 (Table 9): Matched-scale homogeneous-view and homogeneous-expert controls, widened equal-parameter ablations, no-shared-expert variant, and β,γ,λ\beta, \gamma, \lambda loss ablations across all 24 public benchmarks.
  • Section 4.1 & Appendix B.6 (Table 10): Comparison with FreTS, TimeMixer, FEDformer, and Autoformer across all 48 public benchmarks.
  • Section 4.3 & Appendix B.7 (Figure 13 and Table 11): History-only channel-branch covariate forecasting gain (Section 4.3) and test-time future-covariate noise perturbations (s=0.1,0.3,0.5,1.0s = 0.1, 0.3, 0.5, 1.0; Appendix B.7) across all 24 public benchmarks.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.