acceptodds
Withdrawn from ICLR 2027

AnchorScope: Learning Textual Forecast Revisions at the Right Temporal Scale

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Forecasting systems often begin with a prediction from numerical history and then revise it when contemporaneous reports reveal conditions that the observed trajectory has not yet fully expressed. Text is useful in this role when it is correctly paired in time, relevant to the current numerical dynamics, and prevented from overwhelming the numerical evidence. We formulate text-aware forecasting as forecast revision: numerical history establishes a reference from which the horizon is predicted, and paired text proposes a limited update to that reference. AnchorScope distils document histories into semantic candidates, routes them to ordered adaptive frequency bands, and calibrates each revision with bounded amplitude, phase, and residual components. Under a row-level causal availability rule for Climate, aligned text outperforms no text, shuffled text, and random text. Controlled retraining holds encoders, text modulators, budgets, and seeds fixed: anchored adaptive bands outperform fixed, anchor-free, and global allocation rules on four datasets, and configured bounds outperform a threefold relaxation. Broad comparisons on nine Time-MMD datasets show that these revisions retain competitive forecasting performance. AnchorScope learns where a textual revision belongs in the numerical spectrum and how large that revision should be.

Primary area:
learning on time series and dynamical systems
Keywords:
Multimodal time series forecasting, textual context, spectral forecasting, frequency-domain learning, causal text availability

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_vsp4OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes AnchorScope for time series forecasting with textual context. Its main contribution is a frequency-domain revision mechanism that allows text to make bounded adjustments to a reference established from numerical history. The motivation is that text may reveal information not yet fully reflected in numerical observations. For example, historical temperatures may capture existing trends, while contemporaneous reports may describe emerging unusual weather conditions. To incorporate this complementary information, AnchorScope encodes numerical history into a spectral reference and distills textual information into semantic candidates. These candidates are then routed to adaptive frequency bands, where gated amplitude, phase, and residual adjustments determine where and how strongly text revises the numerical representation before the final forecast is generated.

Critical strengths

The paper clearly frames text-assisted forecasting as a revision to a numerical forecast and identifies two central questions: which temporal scales should be revised, and how large should the revisions be? This formulation provides a coherent motivation for the proposed architecture.

The evaluation covers nine domains in Time-MMD, with four forecasting horizons per domain. This breadth allows the method to be assessed across different Time-MMD forecasting domains.

Critical weaknesses
  1. limited novelty:The idea of using numerical history as a reference and text to guide revisions has clear precedents in ChronoSteer, PostTime, and REST-TS. None of these works is discussed or evaluated, weakening the paper’s novelty claim. The contribution appears to lie mainly in combining spectral modeling, attention routing, and gated modulation. The authors should clarify what this combination adds beyond prior revision methods and support the claimed advances with direct comparisons.

2.The motivation for spectral correction remains underdeveloped. While different events may affect different temporal scales, this intuition alone does not explain why text-driven corrections are particularly suited to frequency-specific amplitude, phase, and residual modulation.

3.weak evidence for learned gating: The necessity of learned gating remains unclear. In Table 8, fixing the gate at 1 increases average MSE and MAE by only 0.02% and 0.03%, respectively. These results provide little evidence that adaptive gating offers a meaningful advantage over a constant gate.

4.Small gains lack uncertainty estimates. Some improvements are modest: Table 3 reduces MSE by approximately 0.95% on Climate and 0.25% on Security relative to numerical-only forecasting. Without repeated runs and uncertainty estimates, it is unclear whether these differences exceed training variability.

5.The main and supplementary tables are inconsistent. For TaTS-iTransformer on Agriculture, Table 10 reports MSEs of 0.423, 0.433, 0.520, and 0.599, averaging 0.49375, while Table 2 reports 0.293. On Climate, the corresponding average is 1.0925, while Table 2 reports 0.981. etc.

6.All experiments use Time-MMD. Its nine domains provide useful coverage, but evidence from an independent benchmark is missing. The strict document-availability study is also limited to Climate, leaving its applicability to other domains unresolved.

Questions for authors

See Critical Weaknesses as is.

Reviewer_z4KiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose an architecture that leverages contemporaneous semantic information to refine representations learned from historical numerical data. The method employs an anchored revision point and a controlled revision magnitude to improve forecasting accuracy. Experimental results suggest some potential across various forecasting tasks.

Critical strengths
  1. This work shows a reasonable idea of leveraging specific semantic information to refine numerical representations, emphasizing the location and range of the revision rather than simply relying on global context.
Critical weaknesses
  1. Several key mechanisms and mathematical notations in Eqs. (3)–(5) require further clarification, along with stronger justification for the corresponding design choices, to enable a proper assessment of the proposed method's soundness and novelty.
  2. As a crucial part of the proposed method, the way the semantic candidates are built and the underlying design logic are unclear. For instance, does it explicitly include actual timestamps during construction? How does the temporal semantic adapter work? Why do you need to map the text embeddings to complex tokens with an imaginary part?
Questions for authors
  1. Please elaborate more on how equation (3) is derived, the meaning of the descriptor u_f, and why this way is better for modeling the coefficient W.
  2. Please elaborate more on how the bounds shown in equation (4) are derived.
  3. Please clarify the definition of B_k tilde and the physical meaning of (B_k tilde - B_k) in equation (5)
  4. Do the semantic candidates explicitly include their actual timestamps during construction? How does the temporal semantic adapter work? What is the rationale for mapping the text embeddings to complex tokens with an imaginary part?
Reviewer_jdFYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AnchorScope treats textual context as a revision to a numerical forecasting representation. A numerical spectral encoder forms the reference, a frozen language encoder plus temporal adapter produces semantic candidates, and anchored adaptive bands route textual information to different frequency ranges. Gated amplitude, phase, and residual coefficients modulate band representations before horizon-conditioned decoding. The evaluation combines nine Time-MMD domains, a document-availability-controlled Climate study, and matched retraining for allocation and coefficient bounds. The approach is well motivated, but inconsistent aggregate results and incomplete experimental specification currently prevent a firm assessment of its comparative advantage.

Critical strengths

S1. Clear modeling question The separation between where a text update enters the spectrum and how strongly it acts is intuitive and leads to concrete structural controls.

S2. Matched retraining and temporal controls Table 5 holds encoders, modulators, budgets, and seeds fixed while changing allocation or bounds. The Climate study explicitly restricts document availability and compares aligned, removed, shuffled, and random text. These are more informative than component removal alone.

S3. Broad evaluation and candid diagnostics The study covers nine domains and four horizons, includes numerical/plugin baselines, and reports cases where band masking improves performance. Appendix A also honestly distinguishes coefficient bounds from bounds on the final normalized update.

Critical weaknesses

W1. Aggregate comparisons do not match the detailed results The paper states that domain scores are arithmetic means across four horizons. In Table 10, TaTS-iTransformer Agriculture MSE values 0.423, 0.433, 0.520, and 0.599 average to 0.49375, not Table 2’s 0.293. Traffic values 0.233, 0.215, 0.229, and 0.272 average to 0.23725, not 0.191. These discrepancies exceed rounding error and affect close-baseline comparisons. Corrected per-horizon and aggregate tables, with a common run provenance, are necessary.

W2. Bounded coefficients do not ensure a small or zero revision Appendix A correctly acknowledges that the final correction is not pointwise bounded. More specifically, Eq. (22) multiplies an unconstrained residual proposal by β and then applies layer normalization. At γ=0, the result is LN(B), generally not B, so Eq. (23) can still alter the numerical reference. The coefficients therefore provide a regularization parameterization, not a guarantee that text cannot overwhelm the reference. A matched normalization-preserving no-text control and an identity-preserving gated-update variant would clarify which part of the benefit is genuinely attributable to controlled textual revision.

W3. Reproducibility and uncertainty are insufficiently specified The PDF specifies tensor operations but does not provide a complete concrete configuration: the frozen language encoder checkpoint, K/M/d, adapter/decoder architecture, optimizer and schedule, split sizes, and number of seeds are not consolidated or fully specified. Tables 4–5 give point estimates despite several small differences. The linked anonymous artifact could not be accessed in this assessment, so I cannot verify whether it fills these gaps. Multi-seed paired estimates on the central controls would clarify reliability.

W4. Mechanistic and real-time claims need narrower scope The explicit publication-date reconstruction is demonstrated for Climate, whereas the nine-domain benchmark follows the supplied Time-MMD pairing; date markers alone are not proof of publication-time availability. Also, Table 8’s mean gate is 0.938 and removing it changes MSE by only 0.02%, while low/mid-band masking improves 21/36 and 20/36 settings. The paper should distinguish the supported benefit of the overall architecture from stronger claims that the learned gate reliably suppresses irrelevant text or that all domains meet the same real-time availability guarantee.

Questions for authors

Q1. Which Table 2/Table 10 values come from the intended common protocol, and can all aggregate numbers be regenerated directly from the corresponding per-horizon results?

Q2. How much of the benefit remains with an identity-preserving gate and a no-text control retaining the same numerical transformations and normalization? Can the authors report realized correction-to-reference norm ratios?

Q3. Please provide the complete configuration and paired multi-seed uncertainty for the Climate and structural-control experiments, including how the chronological splits and publication dates are constructed.

Q4. Given the near-saturated gate and masking improvements, when does the gate actually reject irrelevant evidence, and which real-time availability conclusions are supported outside Climate?

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

Related papers

Discussion (0)

Sign in to comment.