acceptodds
Under review as a conference paper at ICLR 2027

Perceive, Route, Modulate: Dynamic Pattern Recalibration for Time Series Forecasting

Abstract

Local temporal patterns in real-world time series continuously shift across regimes. While existing forecasting backbones express input-dependent responses through shared transformations, they lack an explicit mechanism to separate shared representation learning from local feature-gain adaptation. We introduce Dynamic Pattern Recalibration (DPR), a backbone-agnostic mechanism that realizes this separation via token-level modulation. Through a lightweight “Perceive-Route-Modulate” pipeline, DPR computes a soft-routing distribution over a compact learned response basis, generating a context-dependent modulation vector that recalibrates hidden states via a residual Hadamard product. Across ten forecasting backbones, DPR matches or improves the raw backbone’s MSE in 87 of 95 backbone-dataset pairs (91.6%), with larger gains associated with greater local volatility variation. Its minimalist standalone instantiation, DPRNet, achieves strong forecasting performance across 11 benchmarks with low measured computational cost. Our code is available at https://anonymous.4open.science/r/DPRNet.

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

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.50 over 4 reviews, above 99% of the others.
Reviewer_7W1hOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes Dynamic Pattern Recalibration (DPR), a lightweight plug-in module that extracts local temporal context, softly routes each token over a learned response basis, and applies feature-wise residual modulation. DPR is evaluated across multiple forecasting backbones, while DPRNet provides a standalone lightweight instantiation.

Critical strengths

The paper is generally well organized and the experimental presentation is stronger than most submissions in this group. The parameter-matched comparisons against SE, FiLM, gating, and static capacity expansion are useful for isolating the effect of conditional modulation.

Critical weaknesses
  1. The motivation and claimed novelty are not sufficiently separated from generic local conditional feature modulation. The paper motivates DPR by “shifting local dynamics,” but operationally the method performs local-context-conditioned feature-gain modulation through a small response basis. This is a reasonable design, yet it is unclear whether it constitutes a sufficiently new forecasting principle beyond existing local gating, FiLM-style conditioning, and mixture-based modulation. The observed correlation between gain and volatility variation is suggestive rather than a mechanism-level validation.
  2. The empirical comparison is broad in number of backbones but weak in terms of current forecasting competitiveness. Large improvements mainly occur on older or relatively weak backbones such as Informer, whereas gains on stronger recent hosts are generally small. For example, the reported average MSE improvements are only about 1.7% for PatchTST, 3.3% for TimeMixer, 2.9% for TimeFilter, and 1.7% for WPMixer. The standalone comparison also stops largely at 2024–2025 forecasting architectures. By 2026, substantially stronger pretrained/general-purpose forecasters such as Zeus, Sundial, Chronos-2, and multivariate TimesFM-3 are available, making it difficult to assess the practical competitiveness of DPRNet. Zeus, for example, is an ICML 2026 tuning-free foundation model, while Chronos-2 and TimesFM-3 explicitly support multivariate forecasting.
  3. The dataset coverage is not convincing for a method making general MTS forecasting claims. The study excludes widely used high-dimensional traffic benchmarks such as METR-LA, PEMS-BAY, the PeMS sensor benchmarks, and LargeST. This is particularly important because local changes and cross-variable interactions are central in such datasets. LargeST further provides a substantially larger stress test with up to 8,600 sensors, as shown in CPiRi.
  4. The efficiency evidence should more clearly distinguish model size from actual computation. Table 6 is titled as a routing comparison, yet mixes parameter count, accuracy, and measured training/inference time. Parameter count is not a proxy for FLOPs, memory movement, or latency, and the comparison would be substantially clearer with MACs/FLOPs and peak memory under the same hardware and batch settings.
Questions for authors

How does DPR compare with stronger 2025–2026 forecasting models or foundation models under a consistent protocol? Can the authors evaluate DPR on METR-LA, PEMS-BAY, PeMS, or LargeST to demonstrate that the proposed local recalibration remains useful in large, strongly multivariate systems? Can the authors report FLOPs/MACs and peak memory in addition to parameter count and latency?

Reviewer_qP4eOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Dynamic Pattern Recalibration (DPR) for adapting feature responses to changing local dynamics in time series forecasting. DPR uses local context to softly mix a compact set of learned response vectors, then applies their mixture as residual feature-wise gains. This lightweight Perceive-Route-Modulate adapter is placed immediately before the prediction head in the plug-in setting. It matches or improves MSE in 87 of 95 backbone-dataset pairs across ten backbones. A standalone patch-based MLP, DPRNet, also achieves competitive accuracy at low measured computational cost.

Critical strengths
  1. I see the main contribution in the response-basis formulation. Local context selects a mixture of a small shared set of vectors, giving each token a continuous range of feature gains. This adds an explicit adaptation path alongside the host's temporal modeling and differs from direct coefficient prediction or routing through full expert networks.
  2. The small adapter, consistent insertion rule, and documented integration details make DPR straightforward to incorporate into existing forecasters.
  3. Table 2's parameter-matched Local SE, Local FiLM, and gated residual controls are especially helpful for assessing the response-basis formulation. The component ablations and static-capacity results add support.
Critical weaknesses
  1. I read "decoupling" as separating the parameterization of shared transformations and adaptive gains. In the plug-in setting with a linear head, the computation can be written as Wdiag(1+γm(c))hW \mathrm{diag}(1 + \gamma m(c)) h, a structured input-conditioned readout. The gain path still depends on the hidden representation it rescales, and Appendix D.3 explicitly includes this dependence in the full derivative. The main text should tie "decoupling" to this architectural factorization more clearly.
  2. Eqs. (3)-(4) define the routing prototypes as context keys and the response basis vectors as feature-gain directions. What is less clear is the temporal meaning attached to "patterns" in Section 4.5. Figure 6 shows context-dependent mixture weights, but does not associate individual basis vectors with identifiable temporal regimes. The interpretability discussion should distinguish these latent response modes from recognizable temporal structures.
Questions for authors
  1. Does "decoupling" refer specifically to separating shared transformations from context-conditioned gains, given that the gain path still depends on the hidden representation?
  2. Are the routing prototypes and their associated response vectors intended to describe latent response modes, or to correspond to identifiable temporal regimes?
1 comment

Thank you for your detailed response. My concerns have been addressed, and I will keep my positive rating.

Reviewer_z5PHOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a lightweight, backbone-agnostic mechanism DPR to recalibrate representation for shifting local dynamics across time. Subsequently, it also develops a standalone model DPRNet that fully embeds the DPR mechanism. Extensive experiments show that DPR generally improves the performance of forecasting backbones across various datasets. Notably, DPR tends to achieve larger gains on datasets with greater local volatility variation.

Critical strengths
  1. This paper identifies a critical research gap in current literature, the lack of an explicit mechanism for time series forecasting backbones to handle regime shifts.

  2. The proposed DPR mechanism is backbone-agnostic, making it highly compatible with a wide range of advanced architectures.

  3. The results using DPR recalibration look encouraging across forecasting backbones and datasets, and also show its potential on high-volatile, non-stationary time series.

Critical weaknesses
  1. This paper motivates DPR as a solution for regime shifts. However, the datasets mainly cover local scale changes and volatility clustering, which are not equivalent to true regime shifts.

  2. The insights regarding local volatility is interesting, but using dataset-level volatility-of-volatility as a proxy for local heterogeneity is not fully convincing, especially without a specific justification for its applicability to general time series. The paper would benefit from promoting the introduction to this metric from the appendix into the main text.

  3. Table 1 is hard to follow. Almost all numbers are bold and underlined. The caption does not explain the meaning of the bold and underlined text, even though this is clarified in subsequent tables.

Questions for authors
  1. Why is the plug-in DPR applied immediately before the final prediction head rather than the earlier layers? Does its effectiveness shown in results suggest that the representation learned by early layers of the backbones is independent or invariant to regime shifts?

  2. Given the relatively small receptive field of the depthwise convolutions in the perceive stage, can DPR distinguish genuine sudden shifts from anomalies? If the local changes are anomalies, can the approach maintain its performance?

  3. Datasets in this paper are mainly dominated by short-range regime shifts. DPR’s mechanism is good at capturing such characteristics. However, long-range shifts are also common in real-world time series (e.g., finance, climate, energy), and remain an open challenge. Can DPR handle, or be adapted to, such scenarios with simple modifications, like larger convolution kernels and multiscale responses?

  4. How is the boundary between “high-heterogeneity” and “low-heterogeneity” determined using volatility-of-volatility, by a threshold or any criterion? Is the applied threshold or criterion a standard practice? Perhaps a high resulting value, like 1.75, may not fully justify the claimed “high heterogeneity”.

Reviewer_d3mmOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a novel lighweight adapter, dubbed DPR, that rescales the late stage hidden states token by token. It does this by soft-routing the local context over a learned orthonormal basis and applies the result as a residual Hadamard operation. The resulting adapter is evaluated by plugging it in across several forecasting backbones and a wide variety of datasets.

Overall: I think the paper is introduces a simple adapter and showcases it benefits over multiple backbone architectures and datasets. The novelty is limited and its effectiveness claims need to be further understand and refined.

Critical strengths

(a) The lighweight adapter is versatile and can plug into multiple backbones seamlessly and without much work. (b) Experiments show a boost on most of the datasets and backbones and especially the largest gains on non-stationary data.

Critical weaknesses

(i) The novelty is a tad limited in the sense it is essentially similar to input-conditioned feature gating (FILM/soft MoE).
(ii) The experiments show the MoE comparison shows soft-routing beats discrete top-K MoE but not that routing is needed. The baseline lacks a no direct-gate variant or no FILM or gated-FFN baseline. (iii) The gains are typically marginal on most datasets and potentially the largest wins might be due to tiny datasets and/or poorly trained baselines.

Questions for authors
  1. Can we get clarity on what addtional capability is unlocked by DPR that a gated FFN or a FILM layer cannot while keeping same local features and parameter count?

  2. Are the gains significant for ILI and COVID19?


I updated the review based on the discussion with the authors to reflect my new understanding of the results.

Comments on the paper

1 comment

Thank you for the substantial revision. The additional experiments and clarifications have addressed most of my initial concerns, particularly regarding the distinct role of DPR, its effectiveness across forecasting models, and its computational cost. Overall, the revised manuscript is considerably stronger.

I nevertheless see three remaining limitations:

  1. Limited gains on strong backbones. The improvements on recent forecasting models, such as PatchTST, TimeMixer, TimeFilter, and WPMixer, remain relatively modest. The correlation with volatility-of-volatility is interesting, but does not directly establish the proposed mechanism's effectiveness in capturing local dynamic transitions.

  2. Limited cross-variable modeling. DPR primarily performs token-wise feature recalibration. In channel-independent settings, it does not explicitly capture changing dependencies or interactions across variables. The PeMS experiments demonstrate scalability to high-dimensional data, but not dynamic cross-variable modeling.

  3. Scope of the adaptation mechanism. DPR adjusts feature responses within existing representations, rather than explicitly determining when and between which variables information should be exchanged. Extending conditional routing to cross-variable interactions would be an interesting direction for future work.

Overall, these limitations do not outweigh the improvements made in the revision. The responses have resolved most of my initial concerns, and I am therefore inclined to raise my rating from 2 to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.