acceptodds
Withdrawn from ICLR 2027

PRISM: A Training-Free Visual Forecaster for Time Series

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-based time series forecasting provides a promising route for reusing pretrained visual models for temporal modeling. Existing approaches have explored increasingly rich visual representations, multimodal augmentation, learned temporal-visual alignment, and adaptation of pretrained vision models. However, input-format compatibility does not necessarily ensure structural compatibility with pretrained visual representations. A key issue therefore remains underexplored: how to organize complementary temporal factors in a way that is compatible with the model-specific geometry learned during visual pretraining? We term the resulting mismatch between temporal factor organization and the pretrained RGB geometry Time-Series-to-RGB (TS2RGB) misalignment. To address this problem, we propose Prism, a training-free forecasting model built on a frozen MAE encoder-decoder. Temporal Chromatic Rendering (TCR) organizes complementary temporal structures along mutually orthogonal RGB directions derived from the spatially constant color sensitivity of the frozen MAE patch-embedding layer, while Topology-Guided Visual Token Propagation (TVTP) injects dynamic cross-variable dependencies through spatially corresponding latent tokens. Prism introduces no additional trainable parameters and performs forecasting without downstream weight updates. Across eight multivariate benchmarks and four forecasting horizons, Prism ranks first in 43/64 dataset-horizon-metric comparisons and 12/16 dataset-metric averages, with real-token analyses showing 63.5% and 14.9% reductions in cross-component coupling at the patch-embedding interface and final encoder representation, respectively.

Primary area:
learning on time series and dynamical systems
Keywords:
Vision-Based Time Series Forecasting, Zero-Shot Forecasting, Temporal-Visual Alignment, Multivariate Time Series

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_b9RTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper deals with the time series forecasting problem with visual foundation models. It presents PRISM, a training-free model on top of the frozen MAE encoder-decoder.
Two modules are designed: Temporal Chromatic Rendering (TCR) to organize complementary information and RGB directions; Topology-Guided Visual Token Propagation (TVTP) to inject dynamic cross-variable dependencies throught spatilly corresponding latent tokens.

Experiments are conducted on eight standard time-series datasets. Comparisons are made with other pretrained forecasting models such as Timer-XL, VisionTS, etc. Results show that the proposed method beats or in-par-with baseline methods and VisionTS. Albation study and efficiency analysis are also provided.

Critical strengths

This paper raised the Time-Series-to-RGB (TS2RGB) misalignment and proposed a method called PRISM to deal with the misalignement.

  • Temporal Chromatic Rendering (TCR) tries to link the temporal information with the MAE patch-embedding coordinate space, making the time-series data more compatible with Visiton Foundation Model.
  • Temporal-guided Visual Token Propagation (TVTP) explores and reweights the variables relationship.

The method is training-free, without relying on downstream finetuing.

Experiments show that PRISM significantly outperforms existing non-vision baselines, and slightly outperforms VisionTS (appendix).

  • On 12 out of 16 settings (horizon x datasets), PRISM outperforms existing non-vision baselines.
  • In Table 7, PRISM slightly outperforms VisionTS.
Critical weaknesses

Although the paper proposes PRISM as a practical and relatively novel method to solve TS2RGB misalignment. The experiments do not strongly support the usefulness, effectiveness, and efficiency of the proposed method.

  • This paper mainly tackles the problem from Vision Foundation Models, i.e., TS2RGB misalignment. However, the main results Table 1 compares only with non-vision models. This is both unfair and also less convincing.
  • Continue above point, vision-based models are the main comparison baselines in the context of vision-based forecaster. But VisionTS is only compared in the appendix, and the proposed method is only marginal better with sophiscated design.
  • From Table 2, it seems that the w/o Both baseline is already qutie strong. Adding the designed TCR and TVTP introduce marginal improvement. Moreover, this result is only on Horizon=720, while main results are average of all horizons.
  • In Table 3, the training-free method PRISM is faster than other baselines, but has much higher peak memory, leaving the method less practical in deployment. E.g., it costs at least three times Peak Mem. than baseline.
  • Moreover, no comparison with the efficiency of VisionTS.
  • Other vision-based forecaster should also be compared.
Questions for authors

Except the baselines in Table 1, what are the other vision-based methods? Why not compare with them?

Why adding TCR and TVTP introduce marginal or no gain in Table 2?

Why VisionTS is only put in the appendix?

Other mentioned points in the above weaknesses.

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

This paper proposes PRISM, a training-free framework that leverages pre-trained visual/vision-language backbones for time series forecasting. Building upon prior vision-based time series forecasting methods, the primary contributions of this work lie in the input stage: (1) the proposed TCR module, which decomposes and projects different components of the time series onto orthogonal coordinate axes to serve as an input-level temporal augmentation; and (2) an explicit similarity-based mechanism to capture inter-variable correlations before feeding the representations into the visual backbone.

Critical strengths
  1. Training-Free Efficiency: The proposed framework operates in a training-free manner, which significantly reduces computational overhead and avoids expensive task-specific fine-tuning.
  2. Enhanced Input Information via TCR: Compared to prior approaches that naively convert raw time series into images, the TCR module incorporates richer temporal structural priors at the input stage, offering a degree of structural interpretability on the input side.
  3. Explicit Inter-Variable Correlation Modeling: Previous vision-based forecasting methods often handle multivariate dependencies ambiguously. Introducing a classical similarity-based extraction method is a reasonable step forward in explicitly modeling inter-variable relationships.
Critical weaknesses
  1. Lack of Fundamental Soundness, Interpretability, and Controllability: Because the visual/VLM backbone remains completely frozen, the model does not genuinely understand temporal dynamics. Instead, the forecasting task is essentially reformulated as a heuristic image generation/completion task. Relying entirely on the emergent capabilities of a pre-trained visual backbone makes the predictions lack theoretical soundness and interpretability, resulting in very low controllability over the forecasting outputs.
  2. High Sensitivity to Data Quality and Periodicity: The effectiveness of the TCR module appears heavily dependent on the presence of clear periodicity and clean temporal patterns. For high-volatility time series or non-stationary data lacking obvious periodic components, the orthogonal projection in TCR is likely to break down and fail to organize meaningful visual representations.
  3. Insufficient and Coarse Experimental Evaluation: The empirical validation is overly simplistic. Specifically, the fine-grained ablation studies on the core sub-components and the analysis of different visual backbones are severely lacking, making it difficult to verify the true source of the performance gains.
  4. Unverified Scalability on High-Dimensional Datasets: The experiments are restricted to small-scale benchmarks with very few variables. Real-world multivariate time series benchmarks (e.g., Traffic, Electricity, or sensor networks) often contain hundreds or thousands of variables/nodes. Under such high-dimensional settings, pairwise similarity extraction and visual representation construction may suffer from severe memory bottlenecks, quadratic computational complexity, or visual resolution degradation. This lack of evaluation severely limits the practical scalability and generalizability of the method.
  5. Subpar Presentation and Uninformative Illustrations: The overall presentation leaves considerable room for improvement. Specifically, the readability of the main text—particularly in the Methodology section—is poor, making the technical pipeline difficult to follow precisely. Furthermore, the illustrative figures in the main paper appear to be directly generated by AI tools rather than carefully crafted technical diagrams; they lack rigor and fail to clearly convey the concrete mathematical operations and architectural details.
Questions for authors
  1. Clarification on the "Training-Free" Claim: In this "training-free" framework, are there any learnable parameters, projection layers, or hyperparameters that require tuning on the target dataset? Please explicitly clarify which parts are strictly zero-shot/frozen and which parts require optimization or heuristic tuning.
  2. Fine-Grained Ablation Studies on TCR and Similarity Extraction: The current ablation studies are too coarse (mostly removing entire modules) and fail to isolate the actual contributing factors:
    • Regarding TCR: What is the individual contribution of each decomposed component projected onto the three orthogonal axes?
    • Regarding Inter-Variable Similarity: Does the proposed similarity metric genuinely capture meaningful dependencies? If you replace it with a random similarity matrix or alternative distance metrics, does the performance degrade significantly?
  3. Ablation on Visual Backbones: The paper lacks an analysis of how different visual/VLM backbones impact performance. Have you evaluated PRISM across backbones of varying architectures and parameter scales?
  4. Justification of Temporal Understanding vs. Visual Heuristics: Since the visual backbone is not trained on temporal data, how can you justify that the model is learning temporal forecasting rather than merely performing pattern matching in pixel space?
  5. Robustness to Axis and Variable Permutations:
    • The TCR module maps three temporal representations onto three orthogonal axes in a fixed order. Is the model robust if the mapping order of these three representations/axes is permuted?
    • Similarly, does permuting the order of input variables affect the visual representation and final forecasting performance? Please provide empirical results on permutation robustness.
Reviewer_sAFHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a training-free time-series forecasting method built on a frozen MAE encoder–decoder. Its two components are: (1) Temporal Chromatic Rendering (TCR), which maps temporal level, within-period differences, and inter-period differences onto RGB directions derived from the MAE patch-embedding weights, and (2) Topology-Guided Visual Token Propagation (TVTP), which performs correlation-based propagation between spatially corresponding tokens across variables. The paper evaluates PRISM on eight multivariate benchmarks and provides additional ablations, representation analyses, and configuration controls.

The approach is lightweight and reasonably motivated, and the reported results are promising. However, I am not yet convinced that the experiments establish the central claim about pretrained RGB geometry, or that the improvements over the closest predecessor are sufficiently substantial and reliably isolated.

Critical strengths
  1. Both components reuse a frozen visual backbone without introducing additional trainable parameters. The coordinate construction and latent propagation are straightforward and potentially useful for efficient adaptation of pretrained models.
  2. The paper explicitly distinguishes orthogonality under the spatially averaged patch-embedding operator from orthogonality throughout the Transformer. This qualification is important and avoids overstating the analytical result.
  3. The evaluation includes eight datasets, four horizons, matched grayscale controls, fixed structural configurations, context-length controls, and depth-wise representation analyses.
Critical weaknesses
  1. The proposed coordinate construction guarantees orthogonal responses under the spatially averaged patch-embedding operator. However, actual forecasting inputs contain spatially varying temporal patterns, so this property alone does not establish that the resulting directions are better suited to forecasting.
  2. According to Table 5, 17 of the 32 selected configurations disable both phase and cycle augmentation, including every setting on ETTm1, ETTm2, and SDWPF. These settings retain the backbone-aligned level representation but do not use the complementary temporal factors emphasized in the motivation.
  3. The broad baseline comparison is favorable, but the matched VisionTS comparison in Table 7 is more informative about PRISM’s incremental contribution. The dataset-average MSE improvements range from approximately 0.4% to 3.0%, while several MAE improvements are particularly small. The matched-context experiment reports average gains of only 0.62% in MSE and 0.65% in MAE.
  4. The authors transparently define zero-shot forecasting as inference without downstream weight updates. Nevertheless, the method uses target-dataset validation labels to select numerous inference settings. The search space, number of trials, selection cost, and comparable tuning budgets for baselines are not adequately documented. The fixed-configuration experiment helps, but it retains selected context lengths and does not explain how the shared structural configuration was chosen.
Questions for authors

See Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.