acceptodds
Withdrawn from ICLR 2027

JacoWaveNet: Wavelet Residual Pyramid with Jacobi Polynomial Spectral Mapping for Heterogeneous Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time series forecasting benefits from modeling long-range dependencies and cross-scale relations, yet existing approaches often suffer from feature entanglement across scales and insufficient modeling of negative temporal dependencies, limiting their ability to capture heterogeneous temporal interactions. We propose JacoWaveNet, a dual-subband wavelet residual pyramid framework that introduces structured multi-scale representations by separating approximation and detail components, reducing redundant low-frequency semantics and emphasizing scale-wise differential dynamics. Building on this decoupled representation space, we develop a Jacobi Mapping (JM) module based on Jacobi polynomial spectral modeling, which jointly captures global smooth trends and local negative dependencies via polynomial filtering, and progressively models nonlinear temporal relations through Jacobi polynomial message passing. Experiments on seven benchmarks demonstrate consistent improvements over strong baselines, especially for medium and long horizon forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Jacobi polynomial, Wavelet decomposition

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4Y6oOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

In order to model long-range dependencies via multi-scale while preserving the scale-specific semantics, the paper proposes JacoWaveNet, which uses wavelet residual pyramid. The method polish existing ones to solve cross-scale feature entanglement and unsigned temporal aggregation. Experiments show competitive performance of the proposed method.

Critical strengths
  1. The methodology is based on solid theories.
Critical weaknesses
  1. Although the motivation looks somewhat reasonable, it is too abstract and hard to prove. For example:
    • Why repeatly re-encoded smooth trends will obscure scale-specific signals? As long as smooth trends are not fake signals, re-encoding them will not introduce unrelated noise, and there's no evidence provided in this paper that can prove this assertion.
    • Why signed aggregation is definitely better than unsigned one? Theoretically, the signed one indeed distinguishes positive and negative correlations, but can models actually leverage negative correlations instead of being misled?
  2. Method novelty is incremental. The individual ingredients are all established, so the contribution is limited to their specific combination and ordering. Also, the model's small capacity makes attribution of performance gains uncertain.
Questions for authors

Why the proposed method is based on DWT instead of other multi-scale techniques, like Pathformer[1] or Pyraformer[2]?

[1] P. Chen et al., “Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting,” presented at the The Twelfth International Conference on Learning Representations, 2024.

[2] S. Liu et al., “Pyraformer: Low-Complexity Pyramidal Attention for Long-Range Time Series Modeling and Forecasting,” presented at the International Conference on Learning Representations, 2022.

Reviewer_SWgyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

JacoWaveNet runs a per-variable multi-level DWT, keeps each finer coefficient as its residual over the lifted coarser one (for approximation and detail separately), and passes each residual level through a Jacobi Mapping (JM) module. JM computes one signed tanh-score temporal aggregation and then applies an element-wise Jacobi polynomial to the min-max normalized message. The model has hidden dimension 16, is trained for 10 epochs, and is evaluated on seven long-horizon benchmarks against eight baselines, with a five-seed rerun on three datasets.

Critical strengths
  1. The architecture is specified in full and the paper separates what each stage is meant to do, including an honest statement that the Jacobi order is not a temporal-hop parameter.
  2. The paper includes a five-seed rerun with confidence intervals, sensitivity tests for the polynomial basis and the wavelet, and a limitations section.
  3. The model is small and the appendix reports parameter count, latency and memory.
Critical weaknesses
  1. The Jacobi part, which gives the paper its name, does not matter. Table 3 compares equal-order bases, and a plain monomial basis (Weather MSE 0.157) and Chebyshev-T (0.156) beat Jacobi (0.160) on Weather; on ETTm1 Jacobi is 0.306 against 0.307 for Legendre. The authors say the close results are "expected because they span closely related low-dimensional polynomial spaces", which concedes that any polynomial would do. What remains is an element-wise learned polynomial on a min-max normalized message, i.e., a univariate nonlinearity of the kind KAN-style models already use. The "w/o Jacobi" ablation is described only as "replacing JM", without saying what replaces it.

  2. Table 1 is not a controlled comparison, and its margins are inside seed noise. JacoWaveNet runs with an input length of 512 (App. B), but the paper does not give the input length of any baseline, and says that some baseline entries are re-run while "remaining comparison values are taken from the corresponding published reports" without marking which. The JacoWaveNet column is a single seed (2021). The only multi-seed check covers three datasets at H = 96 and shows standard deviations of 0.002 to 0.008 (Table 5). Against that, ETTh2 is a tie with TimeMosaic (0.349 vs 0.349), ECL is a tie with TQNet (0.165), and the model loses on Traffic to iTransformer (0.397) and SOFTS (0.409) against 0.413. Training is 10 epochs with early-stopping patience 10 (Table 6), so early stopping can never act. The external AQWan dataset (Table 13) is reported for JacoWaveNet alone, with no baseline.

  3. The one component with a visible effect, the wavelet residual pyramid, is known, and is not compared with its predecessors. Decomposing a series with wavelets and forecasting the subbands is the subject of WaveForM, WFTNet, SPTGNN and Sonnet, which the paper cites; none of them, nor TimeMixer or DLinear, appears in Table 1. The ablation of the pyramid and of the negative-score branch uses two datasets (Weather and ETTm2) and one seed, and the effect of removing the negative branch on Weather (0.233 to 0.236) equals the seed standard deviation on that dataset (0.003).

Questions for authors

See weaknesses.

Reviewer_4xLTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces JacoWaveNet, a time-series forecasting architecture motivated by cross-scale redundancy in existing multi-scale temporal representations and the need to better capture heterogeneous, signed temporal dependencies. Rather than directly mixing information across temporal scales, JacoWaveNet decomposes the input into wavelet approximation and detail components, constructing separate residual pyramids that emphasize scale-specific information relative to coarser representations. Representations are then processed by the Jacobi mapping module that aggregates positive and negative temporal dependencies separately before applying an element-wise Jacobi polynomial expansion to model nonlinear relationships. Processed representations are then reconstructed through inverse wavelet transforms and mapped to future predictions through a forecasting head. Experiments on several long-term forecasting benchmarks show that this design improves forecasting accuracy, achieving the lowest reported average MSE on 6 of the 7 datasets.

Critical strengths
  1. The proposed architecture is coherent and technically well-defined.

    The paper clearly specifies its dual-subband wavelet residual construction, signed temporal aggregation, and Jacobi polynomial mapping. These components have distinct modeling roles, providing a structured approach to heterogeneous temporal forecasting.

  2. The experiments provide targeted evidence supporting the architectural motivation.

    Beyond standard forecasting benchmarks, the paper includes component ablations and synthetic heterogeneity experiments. In particular, the full model's advantage over its ablated variants generally increases with temporal heterogeneity, providing evidence that the proposed design is beneficial under the conditions it targets.

Critical weaknesses

1. The claimed reduction in cross-scale redundancy is insufficiently justified.

Under exact wavelet reconstruction, Eqs. (7)–(12) imply RlA=LiftD(Cl+1D)R_l^A = Lift_D(C_{l+1}^D). For the default two-level decomposition, R1A=LiftD(R2D)R_1^A = Lift_D(R_2^D), meaning that the finer approximation residual is completely determined by detail coefficients already retained in the other branch. Thus, although residualization removes repeated coarse-scale information within the approximation hierarchy, it introduces deterministic overlap between the two residual pyramids. Furthermore, Table 11 primarily demonstrates the expected decorrelation between wavelet approximation and detail contributions rather than reduced redundancy across the complete representation. The distinct representational benefit of the dual residual construction therefore remains unclear.

2. The temporal relation mechanism imposes restrictive dependencies that are not sufficiently examined.

Equation (18) constructs symmetric affinities using S=tanh(LN(H)LN(H)T)S = tanh(LN(H) LN(H)^T). Although row normalization permits asymmetric magnitudes, the off-diagonal affinities retain reciprocal signs and coupled strengths, restricting independently directed temporal relationships. Additionally, the unscaled dot product may cause the tanhtanh function to saturate even at moderate feature correlations. The paper does not examine whether these restrictions affect its ability to model heterogeneous temporal dependencies.

3. The experiments do not sufficiently substantiate the proposed architectural mechanisms.

Although Table 2 shows that residualization improves forecasting performance, it remains unclear whether these gains result specifically from reduced cross-scale redundancy rather than a different feature representation. Similarly, Table 10 shows that K=1K = 1, which produces only an affine Jacobi response to the normalized messages, achieves the best average MSE among the evaluated orders. These results leave the claimed benefits of residual-based disentanglement and higher-order nonlinear mapping insufficiently established. Additionally, the external-domain results in Table 13 lack baseline comparisons, limiting their ability to establish whether the proposed architecture provides competitive performance beyond the primary benchmarks.

Questions for authors
  1. Under exact wavelet reconstruction, Eqs. (7)–(12) imply RlA=LiftD(Cl+1D)R_l^A = Lift_D(C_{l+1}^D). For the default two-level decomposition, R1A=LiftD(R2D)R_1^A = Lift_D(R_2^D), so the finer approximation residual is determined by detail coefficients already retained in the other branch. What distinct benefit does this duplicated representation provide over processing the corresponding nonredundant wavelet coefficients? Could the authors compare against a conventional wavelet-coefficient representation under matched processing capacity and clarify how their redundancy analysis accounts for this overlap?

  2. Equation (18) constructs symmetric affinities through an unscaled dot product followed by tanhtanh. Row normalization introduces asymmetric magnitudes but retains reciprocal signs and diagonally symmetrizable dependencies, restricting independently directed temporal relations. Moreover, with layer-normalized hidden features, the unscaled scores may cause substantial tanhtanh saturation. Could the authors compare against scaled or asymmetric affinity formulations, or report evidence that these restrictions are beneficial for heterogeneous temporal forecasting?

  3. Although Figure 3 illustrates the greater functional complexity of higher-order Jacobi polynomials, Table 10 shows that K=1K = 1, which yields an affine response to the normalized messages, achieves the lowest average MSE on ETTm2. While higher orders improve certain horizons, their benefits are not consistent across the reported settings. Under what conditions do higher-order terms provide a meaningful advantage over the first-order formulation? In particular, can the authors demonstrate that these gains are associated with stronger localized or heterogeneous temporal dynamics rather than simply additional polynomial flexibility?

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

Related papers

Discussion (0)

Sign in to comment.