acceptodds
Under review as a conference paper at ICLR 2027

ASE: Attribution-Guided Semantic Evolution for Time Series Forecasting

Abstract

External text can supply predictive information absent from numerical histories, but its value depends on how it is represented, combined and deployed. We present Attribution-guided Semantic Evolution (ASE), a modular framework for constructing textual factors, evaluating their conditional contribution, and deploying validation-selected numerical-textual predictors. ASE supports task-specific implementations with lexical measurements, semantic primitives or evolved event/state factors, coupled to suitable numerical experts. Its feedback-based implementation uses conditional refitting diagnostics to guide factor revision and combination search. We evaluate ASE on the four text-associated TESS datasets and dynamic CSI 300 stock ranking. The four forecasting implementations obtain test MSEs of 0.001041, 0.022439, 0.333115 and 1.8150, each below the corresponding published TESS score. On CSI 300, the frozen text-fusion implementation improves mean RankIC by 0.000642 over its numerical expert and by 0.000701 over matched shuffled text, with positive differences in all five seeds. Controlled text perturbations support functional reliance on text correspondence, while incumbent retention prevents regressions in the measured development criterion. Together, the experiments characterize ASE at two levels: forecasting performance of task-specific implementations and the conditional contributions of individual modules.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, text-augmented forecasting, large language models, semantic factor evolution, conditional feature selection, multimodal learning

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

This paper presents Attribution-guided Semantic Evolution (ASE), a modular framework for text-augmented time series forecasting. It consists of three modules: semantic factor construction, conditional combination evaluation, and validation-controlled deployment. Through a predictor-consistent feedback loop, it evaluates the conditional contribution of each textual factor and selects effective predictors. The method is validated on four TESS datasets and a dynamic CSI 300 stock ranking task, with controlled module-level analyses.

Critical strengths

One key strength is the feedback-based semantic evolution mechanism, which links textual factor revision to measurable conditional gains from predictor refitting. This provides a principled way to quantify textual feature utility.

Critical weaknesses
  1. The main TESS benchmarks only compare point estimates without paired statistical tests
  2. The CSI 300 numerical reference confidence interval includes zero.
Questions for authors
  1. Introduction mentions "We study a different feedback question: what information should a forecast model return to the language model that proposes its textual measurements?". Why cannot existing text forecasting methods answer this question, and what is the specific research gap?
  2. The scoring function of combination search in Section 3.2 introduces block standard deviation penalty and subset size penalty. What is the basis for the values of the weight coefficients of the two penalty terms, and has parameter sensitivity analysis been conducted?
  3. Section 3.1 mentions "Six raw factors and eight event–state products are compiled per generation; after five rounds the accumulated library has 70 feature slots.". What is the basis for this hyperparameter setting, and is there a risk of factor redundancy?
  4. The main leaderboard in Chapter 4 Experimental Setup mixes published TESS scores and results of locally adapted models. Given differences in implementation and tuning degree between the two, how to ensure the fairness of direct comparison?
  5. The confidence interval of the numerical reference in the CSI 300 experiment includes zero. Does this affect the statistical robustness of the text fusion gain?
  6. Section 5.3 Module Analysis only selects partial tasks to verify component benefits. Why are systematic cross-task ablation experiments not designed?
  7. Full names should be given for all English abbreviations at their first occurrence. Many English abbreviations such as LLM lack full names. There is a lack of linguistic rigor.
Reviewer_qHLWOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes ASE, a framework for improving textual features for time series forecasting. ASE comprises three interfaces: C1 constructs or modifies textual factors; C2 evaluates their conditional contributions by refitting predictors and searching for effective factor combinations; and C3 either freezes the predictor or deploys it with a gating mechanism selected using the validation set. The author uses Qwen model to generate event and state definitions, constructs textual features through embedding similarity and event–state products, obtains feedback from a Ridge predictor, and retains combinations from the previous iteration. The manuscript reports task-specific forecasting performance and ranking results on four TESS datasets.

Critical strengths

The ASE method is clearly defined and validated. ASE evaluates the contribution of textual features through changes in loss after adding or removing features and refitting the predictor, using the same predictor to provide feedback and make final predictions.

Critical weaknesses
  1. Ablation studies are important. The manuscript currently reports ablation results for only some tasks. More comprehensive ablation studies are needed.

  2. The current experiments provide limited evidence for the stability of performance gains and predictive generalization. For example, the Bitcoin contains only 13 time windows, making it difficult to draw conclusions.

Questions for authors
  1. Which C1/C2/C3 operations were actually used for each of the four main leaderboard results? Which results involved multiple rounds of revision driven by conditional contribution feedback?

  2. Under the same predictor and experimental settings, what specific performance gains does ASE achieve over existing approaches?

Reviewer_JbGROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Attribution-guided Semantic Evolution (ASE), a framework for integrating textual information into time series forecasting. ASE uses LLMs to generate and refine semantic factors, evaluates their contributions through conditional predictor refitting, and selects numerical–textual combinations for forecasting. The authors evaluate the framework on four TESS datasets and a CSI 300 stock-ranking task.

Critical strengths

Relevant research direction. Integrating LLMs with time series forecasting models is an important research topic. The proposed framework explores a potentially useful approach to incorporating textual information into numerical forecasting.

Efforts to control experimental confounding factors. The authors conduct several ablation studies and controlled experiments to isolate the contributions of different components. These efforts are appreciated.

Critical weaknesses

Limited methodological novelty. The proposed framework largely combines existing techniques, including LLM-based factor generation, feature selection, and reflection-based optimization. The paper does not sufficiently demonstrate the independent contribution of its proposed feedback mechanism over existing approaches. Without stronger evidence, ASE appears to be a relatively straightforward combination of established methods.

Inconsistency between the proposed framework and the main experiments. The main experiments evaluate several task-specific text-augmented forecasting systems rather than consistently applying the complete iterative ASE framework across datasets. Consequently, it is unclear to what extent the reported improvements can be attributed to the proposed semantic evolution mechanism.

Insufficiently controlled baseline comparisons. The main leaderboard combines published TESS results with locally adapted baselines, despite differences in model architectures, implementations, and hyperparameter tuning budgets. For example, ASE uses an ensemble of six LightGBM predictors on Electricity, while some baselines employ relatively small neural networks. The improvements may therefore result from stronger numerical backbones or ensemble strategies rather than the proposed textual factor construction method. These comparisons do not provide sufficiently convincing evidence of ASE's superiority.

Concerns about experimental scale and statistical reliability. The Bitcoin dataset contains only 13 test windows, which is insufficient for drawing robust conclusions about generalization. In addition, the limitations concerning timestamp verification and temporal alignment raise concerns about the reliability of the evaluation.

Presentation and writing quality. The manuscript would benefit from substantial editing. Some experimental details, such as random seed configurations, are repeated unnecessarily. Figures 1 and 2 make inefficient use of space, and many numerical results are reported with excessive decimal precision. A more concise and better-organized presentation would improve readability.

Questions for authors

Could the authors expand the experiments and provide more statistically reliable evaluations, particularly on datasets with larger test sets?

Could the authors clarify which experiments implement the complete ASE framework and demonstrate the independent contribution of the proposed semantic evolution mechanism under controlled comparisons?

1 comment

As the authors have not responded to the review, my concerns remain unaddressed. I therefore maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.