acceptodds
Under review as a conference paper at ICLR 2027

GRAFT: Decoupling Dynamic Recovery from Structural Prediction in Time Series Forecasting

Abstract

Modern long-term forecasters are typically trained with global point-wise objectives, whose error reduction is dominated by high-energy, low-frequency periodic structures. This bias favors structural forecasts, while high-utility sparse local dynamics are systematically under-rewarded and risky to optimize. We therefore propose GRAFT, a general forecasting enhancement paradigm that decouples dynamic recovery from the structure-dominated role of a base forecaster. GRAFT treats dynamic recovery as a distinct forecasting target with its own modeling process, while the existing forecaster retains its full prediction as a structural anchor. Specifically, historical context–response pairs provide evidence for recurrent local dynamics, which GRAFT selectively grafts onto the base forecast under phase-, channel-, and confidence-aware constraints. Complementary objectives explicitly supervise local changes and global spectral structure, guiding dynamic recovery beyond point-wise accuracy alone. Across seven benchmarks, two metrics, four forecasting horizons, and eight representative backbones, GRAFT improves over the corresponding backbone in 85.7% of all comparisons, with average MSE/MAE reductions of 3.88%/3.00%. Dynamic-window analyses further show that the gains arise from recovering local dynamics while preserving the stable backbone. These results validate the effectiveness of decoupled dynamic recovery and demonstrate consistent enhancement. More broadly, they suggest a new principle for high-utility forecasting: predict the backbone, then graft the dynamics. Anonymous code is available at: https://anonymous.4open.science/r/GRAFT-5E8F/.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Dynamic Recovery, Forecasting Enhancement

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_nWv5OpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1(−1)
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper proposes GRAFT, a plug-and-play framework that decouples structural forecasting from local dynamic recovery. It retrieves historical context–response patterns and selectively incorporates them into backbone predictions through phase-aware calibration, channel-wise modulation, and confidence gating. Additional objectives supervise local temporal changes and global spectral consistency. Experiments across seven datasets, eight forecasting backbones, and four prediction horizons report consistent improvements in MSE and MAE, supported by ablation studies, comparisons with existing enhancement methods, and dynamic recovery analyses.

Critical strengths
  1. Broad empirical validation: GRAFT is evaluated across seven datasets, eight forecasting backbones, and four prediction horizons, demonstrating relatively consistent improvements. This provides meaningful evidence of its applicability across different forecasting architectures.

  2. Targeted dynamic recovery: The paper addresses the under-recovery of local temporal variations and provides dynamic-region analyses and ablation studies supporting its effectiveness. This offers a practically relevant perspective beyond aggregate forecasting accuracy.

Critical weaknesses
  1. Limited methodological novelty: GRAFT primarily combines established techniques, including retrieval-augmented forecasting, post-hoc prediction correction, confidence gating, and auxiliary temporal losses. Closely related ideas have already been explored in RAFT (Han et al., 2025), PIR (Liu et al., 2025), GTR (Cao et al., 2026), and δ-Adapter (Liang et al., 2026). The proposed framework appears to be an incremental architectural integration rather than a fundamentally new forecasting paradigm. Its distinction from these methods is insufficiently established.

  2. Insufficient justification of dynamic decoupling: Despite claiming to recover local dynamics separately, GRAFT directly scales and adds retrieved full future responses without explicitly separating dynamic components from structural information. Moreover, the positive multiplicative gates constrain the correction direction to follow the retrieved response. The experiments do not convincingly establish that improvements arise specifically from dynamic recovery rather than general retrieval-based correction.

References

[1] Han, S., Lee, S., Cha, M., Arik, S. O., & Yoon, J. (2025). Retrieval Augmented Time Series Forecasting. Proceedings of ICML 2025, 21774–21797. https://proceedings.mlr.press/v267/han25d.html

[2] Liu, Z., Cheng, M., Zhao, G., Yang, J., Liu, Q., & Chen, E. (2025). Improving Time Series Forecasting via Instance-aware Post-hoc Revision. NeurIPS 2025. https://doi.org/10.52202/085713-1196

[3] Cao, F., Dai, L., Han, J., & Xiong, H. (2026). Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval. ICLR 2026. https://arxiv.org/abs/2602.10847

[4] Liang, D., Li, Q., Wang, Y., et al. (2026). The Forecast After the Forecast: A Post-Processing Shift in Time Series. ICLR 2026. https://arxiv.org/abs/2601.20280

Questions for authors
  1. Methodological novelty: What fundamentally distinguishes GRAFT from existing retrieval-augmented and post-hoc correction methods, particularly RAFT, PIR, GTR, and δ-Adapter? Which specific mechanism constitutes a substantive methodological advance beyond their combination?

  2. Validity of dynamic decoupling: Since GRAFT directly adds gated, scaled historical future responses to backbone predictions, how does it explicitly separate local dynamics from structural components? What prevents redundant structural information or inappropriate corrections from being introduced?

  3. Mechanism verification: The dynamic-window analysis demonstrates error reductions in highly variable regions but does not directly verify the proposed grafting mechanism. Can the authors provide quantitative evidence that retrieval relevance, phase alignment, and gate activation specifically drive local dynamic recovery rather than generic prediction correction? What additional evidence distinguishes this mechanism from the residual-MLP control already reported?

2 comments
AuthorsOpenReview ↗

Response to Reviewer nWv5

We thank the reviewer for the assessment and for recognizing GRAFT's empirical validation.


Response to W1

We appreciate the reviewer’s detailed comparison with related forecasting methods.

First of all, the reviewer's selection of closely related works itself do not substantiate the reviewer's characterization of closely related methodological precedents. GTR and δ-Adapter belong only to the broad category of forecasting enhancement, and do not even address the same research problem as GRAFT. GTR addresses the limited modeling of global periodic patterns through global temporal retrieval, whereas δ-Adapter focuses on lightweight input adaptation and output-residual correction. And criticism rests on a category error: component-level overlap, such as the use of retrieval (RAFT, PIR), does not establish that the same forecasting formulation already exists, nor does it imply equivalence in the underlying forecasting problem, modeling objectives, or methodological contributions.

The claim overlooks both conceptual distinctions and experimental evidence.

Across eight backbones, including PMDformer, Dynamic-TMoE, and XLinear, and 56 dataset–backbone settings, GRAFT improves at least one metric in 54/56 cases (96.4%; Table 1). It achieves best or tied-best results in 11/14 dataset–metric comparisons against plug-and-play enhancement methods, including PIR and TimeAlign (Table 2). Thus, component overlap does not make GRAFT an incremental integration.


Response to W2

The critique conflates decoupling forecasting responsibilities with explicit separation of dynamic components, treating clean decomposition as a prerequisite for dynamic recovery. If the reviewer means separating a dominant component and residual (e.g., FFT filtering), this does not cleanly separate structure and dynamics: the residual also contains noise. Moreover, local dynamics are context-conditional and anchored to the structural trajectory; mechanical separation risks losing the context determining when and how they occur. GRAFT instead retains complete historical responses as predictor-independent evidence and infers transferable portions through a dedicated recovery pathway.

The positive-gate concern conflates amplitude gating with direction control. The gate does not generate correction directions; it controls the transfer intensity of retrieved prototypes that already contain upward, downward, and turning-point variations. Negative coefficients could invert historically supported changes (e.g., an upward surge into a downward drop), undermining retrieved evidence and reducing grafting to unconstrained residual correction. Direction and transfer intensity are distinct responsibilities.

Table 3 shows that replacing GRAFT with Naive Prototype or Residual MLP causes consistent degradation, often below the original backbone. Table 4 and Figure 3 further show that gains are not generic: across Q1–Q5 dynamic-window groups, error reductions concentrate in highly dynamic regions (Q3–Q5), including sharp turning points and abrupt peaks and valleys.


Response to Q1

This question overlaps with W1. We refer the reviewer to the methodological distinctions and controlled comparative evidence provided there.


Response to Q2

This question reiterates the concern addressed in W2. Complete historical responses are candidate evolutionary evidence, not values directly added to the backbone forecast. GRAFT filters candidates through shape, differential, and spectral homology retrieval, then controls transfer timing and magnitude through phase anchoring, channel-wise bounds, and confidence gating to suppress incompatible or redundant supplementation. Further details are provided in W2 and the main ablation studies.


Response to Q3

We appreciate the request for further mechanism verification. However, isolating retrieval relevance, phase alignment, and gate activation presupposes independent additive contributions. Dynamic recovery is context-conditional: these mechanisms jointly form a nonlinearly coupled dynamic transfer field (Eq. 4).

Recovery requires similar historical evolution as evidence, compatibility with the current prediction horizon, and compensation when the backbone has not already captured the event. Isolating any one mechanism cannot characterize this coordinated process.

Table 3 and Appendix B.7 provide evidence for the joint mechanism. The Residual MLP control retains the same backbone and retrieved prototype but replaces phase-anchored, confidence-gated transfer with generic residual fitting; its substantial degradation shows that gains depend on coordinated transfer constraints, not generic correction. Section 4.4 shows error reductions concentrated in moderate- and high-dynamic intervals, while Figure 3 visualizes recovery of sharp drops, rebounds, and turning points smoothed out by the backbone.

We thank the reviewer again and hope these clarifications address the concerns.

I appreciate the authors' clarifications. However, my main concerns remain unresolved. The rebuttal does not sufficiently establish substantial methodological novelty beyond existing retrieval-based correction methods. Moreover, the positive-gated addition of complete historical responses does not convincingly demonstrate genuine dynamic decoupling. The existing ablations also fall short of directly validating the claimed mechanism. While the empirical results are encouraging, they do not fully address these fundamental concerns. I therefore maintain my original score.

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

This paper proposes GRAFT, a plug-and-play framework for time series forecasting that decouples backbone prediction from local dynamic recovery. It retrieves similar historical patterns and selectively integrates their future dynamics into backbone predictions through phase-aware calibration and confidence gating. Experiments across multiple datasets and forecasting backbones show consistent improvements in forecasting accuracy, with gains particularly concentrated in highly dynamic regions.

Critical strengths

S1: The authors conduct extensive experiments across multiple datasets and settings, demonstrating that the proposed method outperforms existing state-of-the-art methods. The method is also effective across various backbones and datasets.

S2: The Related Work section clearly explains the differences between the proposed method and existing approaches.

Critical weaknesses

w1: The motivation is unclear, and the paper does not sufficiently establish the necessity of the proposed approach or clearly identify the underlying problem. Regarding Figure 1, could the authors explain why misaligned dynamic recovery might be penalized more severely than overly smooth and conservative predictions?

w2: The novelty is limited. The paper does not introduce a fundamentally new perspective or technique for addressing the problem, and its contributions appear to be primarily a combination of existing ideas.

w3: In Table 2, the authors compare GRAFT with several similar methods but do not explain why GRAFT outperforms them. Are the improvements attributable to the number of parameters or to architectural differences?

w4: Could the authors provide a more in-depth analysis in the ablation studies to better demonstrate the effectiveness of the proposed method?

Questions for authors

Please refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer UXNX

We thank the reviewer for the constructive feedback and for recognizing GRAFT's empirical effectiveness and related-work discussion.


Response to W1

We appreciate the reviewer’s question about GRAFT's motivation and Figure 1. The “misaligned dynamic recovery” raised here is precisely the central problem formally analyzed in our paper.

Point-wise objectives tend to suppress sparse local dynamics across models and datasets (Figure 1). We demonstrate this phenomenon and characterize its optimization incentives. Appendix A.1 provides two results: Lemma A.1 shows that dominant structural components yield greater MSE reductions; Lemma A.2 proves that, for temporally non-overlapping impulse events, a misaligned prediction incurs exactly twice the MSE of completely missing the event (Eq. 19)—the double penalty of missing the true event and generating a false one.

This analysis establishes GRAFT's motivation: conventional point-wise optimization favors dominant structures while potentially imposing greater penalties on inaccurate dynamic recovery. We therefore propose Predict Backbone, Graft Dynamics: the backbone handles stable structural prediction, while a dedicated branch recovers local dynamics from historical response evidence, addressing the trade-off between stable structure and high-penalty dynamic recovery.


Response to W2

We appreciate the reviewer’s attention to GRAFT's methodological novelty. This criticism conflates the reuse of technical components with a lack of novelty. Its contribution is not the individual use of retrieval, gating, or residual addition, but a new division of forecasting responsibilities rooted in optimization asymmetry that existing techniques struggle to address. Rather than jointly fitting structural patterns and sparse dynamics under one objective, GRAFT retains the full backbone prediction as a structural anchor and separately models dynamic recovery using historical context and response evidence.

Controlled experiments support this distinction. With the same backbone and retrieval prototype, replacing controlled grafting with a generic Residual MLP performs worse than GRAFT in all 10 comparisons and degrades the original backbone in 8 (Table 3; Appendix B.7). Replacing complete-response memory with backbone-dependent residuals, while keeping other mechanisms unchanged, worsens 9/10 comparisons (one tie).

These results show that gains cannot be attributed solely to retrieval or generic residual correction. GRAFT decouples structural prediction from history-based dynamic inference, assigning distinct evidence representations, modeling responsibilities, and learning objectives to the two roles.


Response to W3

We thank the reviewer for raising the question of what drives the improvements in Table 2.

The reviewer’s concern that performance gains are driven by parameter count or architectural differences reflects a misinterpretation of the positioning of plug-and-play post-hoc enhancement methods.

Unlike independently designed end-to-end models, GRAFT and the compared methods in Table 2 (PIR, TimeAlign) are lightweight enhancement modules. Their core design principle is to leave the original backbone architecture unchanged and supplement information to improve predictions only through low-cost auxiliary components. Accordingly, parameter count or architectural difference is not the proper evaluation dimension for this category of methods.

For post-hoc enhancement methods, the core metric that truly determines their practical utility is the trade-off between the overhead of functional modules such as memory and retrieval, and the resulting performance improvement relative to the backbone model. We present a dedicated and thorough quantitative analysis of this trade-off in Appendix A.6 (as well as in Table 5 and Table 6). Therefore, our experiments not only validate the effectiveness of the proposed enhancement mechanism, but also fully examine its primary computational trade-offs.


Response to W4

We appreciate the suggestion. Table 3 and Appendix B.7 provide controlled ablations demonstrating component effectiveness and why GRAFT's decoupled design is necessary beyond generic retrieval-based correction.

The key evidence is the Residual MLP control: with the same backbone and retrieved response prototype, replacing controlled dynamic grafting with a generic residual adapter underperforms GRAFT in all 10 comparisons and degrades the original backbone in 8. Thus, additional historical information and an auxiliary correction branch alone are insufficient.

Two further controls support this distinction: replacing complete historical responses with backbone-dependent residuals worsens 9/10 comparisons (one tie), while removing confidence-aware gating or restoration-cooperative optimization consistently degrades performance.


We thank the reviewer again for the thoughtful feedback and hope these clarifications address the concerns.

Thank you for your detailed responses and clarifications. While I acknowledge the empirical effectiveness of GRAFT, I believe my main concerns regarding the motivation, methodological novelty, and sources of performance improvements have not been sufficiently addressed. In particular, the additional explanations and ablation results do not fully clarify how GRAFT differs fundamentally from existing approaches or why its design is necessary. Therefore, I will maintain my original rating.

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

This paper addresses the tendency of time series forecasters to achieve low average error while missing local changes such as sudden spikes and drops. GRAFT freezes the backbone, retrieves similar historical contexts and their future responses, and selectively adds corrections through temporal and channel-wise modulation and gating. The paper reports improvements across several backbones and datasets.

Critical strengths
  1. Focuses on local changes that aggregate forecasting metrics may not adequately capture.
  2. Includes evaluation across multiple backbones, component ablations, and error analysis by local variation intensity.
Critical weaknesses

1.The connection between the stated problem and the proposed solution is not sufficiently established. Better accuracy with retrieval-based correction does not demonstrate that GRAFT resolves the limitations of point-wise training or separates structural prediction from dynamic recovery. The conceptual distinction from existing methods such as PIR is also limited. 2. The method relies on historical responses remaining useful for the current input. There is no direct evaluation of whether the gate prevents harmful corrections when relevant historical examples are unavailable or the relationship between past contexts and future responses changes.

Questions for authors

1.How can we tell whether the gains come from addressing the limitations of point-wise training, beyond simply using additional historical information? 1-1. Why is a separate retrieval and correction branch needed compared with applying the same difference and spectral losses directly to the backbone? 2. When similar historical inputs lead to different future behavior under new conditions, how does GRAFT recognize that the retrieved responses are no longer useful? 2-1. Is there evidence that the gate can suppress harmful corrections in such cases?

2 comments
AuthorsOpenReview ↗

Response to Reviewer TZJN

We thank the reviewer for recognizing the importance of local dynamics and for the constructive assessment.


Response to W1

We agree that improved aggregate accuracy from retrieval alone does not establish dynamic decoupling. However, GRAFT does not claim to “resolve the limitations of point-wise training.” Section 1 and Appendix A.1 show that point-wise objectives favor dominant structure while under-rewarding sparse, uncertain dynamics. GRAFT does not eliminate this limitation; it addresses under-recovery of high-utility dynamic transients. Direct evidence beyond aggregate accuracy is provided in Figure 3 and Table 4, show gains concentrated in highly dynamic regions, not uniformly across forecast points.

Methodologically, GRAFT retains the backbone's complete forecast as a structural anchor and assigns dynamic recovery to a separate, history-grounded process. Decoupling forecasting responsibilities does not mean isolating local dynamics from their structural context. Local events are context-conditional: their occurrence, timing, direction, and magnitude depend on the surrounding temporal evolution. Accordingly, GRAFT (Eqs. 1–4) explicitly formulates dynamic recovery as:

\underbrace{\beta_{\mathrm{gate}}\odot\Lambda_{\mathrm{anchor}}\odot R}_{\mathrm{dynamics}}.$$ The distinction from retrieval-based instance-level revision (PIR) or generic residual correction is not whether history is retrieved, but how structural context is retained for dynamic recovery. A residual corrector targets backbone error: $$E_i=Y_i-\hat{Y}_{\mathrm{base}},\quad g\phi(R_i)\approx E_i,\quad \hat{Y}_{\mathrm{res}}=\hat{Y}_{\mathrm{base}}+g\phi(R_i).$$ Subtracting the backbone forecast removes the structural context governing local events from the correction target. GRAFT retains complete historical responses, using structural anchors and historical patterns for conditional dynamic transfer. Table 3 validates this distinction: with identical backbone and retrieved prototype, GRAFT outperforms the Residual MLP in all 10 dataset–metric comparisons. With a shared PatchTST backbone, GRAFT is best or tied-best in 11/14 comparisons against enhancement methods including PIR (Table 2). --- ### Response to W2 We appreciate this concern. GRAFT does not assume historical responses are always transferable: they are candidate evidence controlled by a confidence-aware gate (Eqs. 4, 11): $$g^{(c)}=[\hat{\ell}^{(c)}\Vert\mathrm{Sim}_{\mathrm{mixed}}^{(c)}\Vert z_{\mathrm{pred}}^{(c)}],\quad \beta_{\mathrm{gate}}^{(c)}=\operatorname{Sigmoid}(\operatorname{MLP}(g^{(c)})).$$ The gate considers retrieval reliability, estimated backbone error, and current prediction state. When evidence is unreliable or correction unnecessary, $\beta_{\mathrm{gate}}\approx0$ suppresses transfer; when evidence is reliable and correction needed, $\beta_{\mathrm{gate}}\approx1$ releases dynamic recovery. History is selectively used, not blindly injected. Because standard benchmarks lack instance-level annotations of historical evidence unreliability; therefore, so the reviewer's proposed direct verification cannot be performed using the available annotations alone. However, Table 3 provides indirect evidence through a controlled ablation: if confidence-aware gating protects against unreliable or unnecessary historical transfer, removing this mechanism should worsen forecasting performance: removing the gate degrades performance across all 10 dataset–metric comparisons, supporting its role. --- ### Response to Q1 & Q1.1 Q1. Extra historical information alone cannot explain the gains. With the same backbone and retrieved prototype, a generic Residual MLP loses to GRAFT in all 10 comparisons and even underperforms the original backbone in 8/10 (Table 3). Together with Figure 3 and Table 4, this supports conditional dynamic transfer rather than retrieval alone. Q1.1. We appreciate the suggestion to apply difference and spectral losses directly to the backbone. Yet supervisory objectives must align with modeling responsibilities. Although the backbone predicts the full trajectory, point-wise optimization tends to favor dominant structural regularities; the grafting branch specifically recovers underrepresented local dynamics. Difference-based supervision encourages dynamic recovery, while spectral supervision constrains the resulting forecast's structure. Applying both directly to the backbone couples these distinct responsibilities in one optimization pathway, potentially causing gradient interference and compromising their respective objectives. The separate branch is therefore a deliberate alignment of roles and supervision, not an arbitrary architectural addition. --- ### Response to Q2 & Q2.1 These questions overlap with W2. We refer to the mechanism, explicit gate supervision, and ablation evidence already detailed there (Eqs. 4, 11; Appendix A.5; Table 3).

Thank you for clarifying the intended meaning of decoupling and the evidence beyond generic residual correction. However, my two main questions remain unresolved. The response does not compare GRAFT with applying the same difference and spectral losses directly to the backbone; gradient interference remains a hypothesis. Likewise, removing the gate on standard benchmarks does not establish that it suppresses harmful corrections when similar historical contexts lead to different future behavior. While I acknowledge the reported empirical gains, these concerns remain, so I maintain my original rating.

Comments on the paper

1 comment

Dear Reviewers,

The authors have replied to your initial comments. Please have a look and engage in the discussion.

Thanks!

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.