acceptodds
Under review as a conference paper at ICLR 2027

EviTime: Evidence-Guided Vision-Language Learning for Multimodal Time Series Forecasting

Abstract

Multimodal time series forecasting aims to exploit complementary numerical, visual, and semantic information for more accurate predictions. However, existing methods often treat temporal observations as equally reliable, rely on limited visual representations to characterize temporal structures, and introduce generic textual descriptions that are weakly aligned with individual forecasting instances. To address these limitations, **EviTime** is proposed as an **Evi**dence-guided vision-language learning framework for multimodal **Time** series forecasting. First, EviTime introduces evidential reasoning into temporal representation learning by combining Dempster-Shafer Theory (DST), nonlinear Basic Probability Assignment (BPA) generation, and Deng entropy to characterize observation reliability and suppress ambiguous information. Second, numerical sequences are further transformed into three-channel visual representations using Gramian Angular Summation Fields (GASF), Gramian Angular Difference Fields (GADF), and amplitude information, enabling a pretrained Vision-Language Model (VLM) to capture complementary structural patterns. Third, hierarchical metadata prompts encode dataset-, task-, and sample-level semantics, while a cross-attention metadata router extracts instance-relevant language information. Temporal, visual, and semantic representations are then aligned and adaptively fused for forecasting. Finally, extensive experiments on multiple benchmarks demonstrate consistent improvements over competitive baselines, with Mean Squared Error (MSE) and Mean Absolute Error (MAE) reduced by 18.4% and 13.6%, respectively, compared with the average performance of 10 state-of-the-art methods. The source code is publicly available at https://anonymous.4open.science/r/EvidentialTime-CD2F.

Primary area:
learning on time series and dynamical systems
Keywords:
Multimodal Time Series Forecasting, Dempster-Shafer Theory, Vision-Language Model, Representation Learning, Evidential Reasoning

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

EviTime is a multimodal long- and short-term forecaster with three branches. (i) TEFE (Sec. 3.2) generates "BPAs" over K = 7 focal elements of a 3-state reliability frame (reliable / locally perturbed / unreliable) with temporal and variable MLPs plus softmax. These are weighted by evidence-consistency/strength terms, combined via a Deng-entropy gate (Eq. 1), and regularized with adversarial, sparsity and orthogonality losses (Eq. 2). They are then added to a linear temporal/variable projection (Eqs. 3–4). (ii) GVFE (Sec. 3.3) turns the series into a 3-channel GASF/GADF/amplitude image, averages it over variables, and encodes it with BLIP-2's vision encoder and Q-Former (Eq. 5). (iii) MLFE (Sec. 3.4) builds dataset-, task- and sample-level metadata prompts encoded by BLIP-2's text side, with a 3-token cross-attention router (Eq. 6). The branches are fused by cross-attention and gating (Eq. 7) and trained with MSE + L_TEFE (Eq. 8). Experiments cover 7 LTSF benchmarks against 10 baselines (Table 1), M4 (Table 2), sensitivity (Fig. 3), ablation (Table 3), noise robustness (Fig. 4), and few-/zero-shot results (Tables 4–5).

Critical strengths
  • [Importance of the research problem] Using vision-language models for multimodal time series forecasting is an active, practically relevant direction.
  • [Importance of the research problem] Instance-specific metadata prompts, computed only from the look-back window (Sec. 3.4.1), target a real weakness of generic text prompts.
  • [Novelty of the approach] Combining evidential representations, three-channel GAF images (GASF, GADF and amplitude) and hierarchical metadata in one VLM-based forecaster is a new combination.
  • [Soundness and feasibility] The module split (TEFE, GVFE, MLFE, fusion; Secs. 3.2–3.5) is clear and feasible to implement. The paper links anonymous code and gives complexity analysis (App. B).
  • [Soundness and feasibility] The amplitude channel sensibly recovers the magnitude information that the arccos mapping loses (Sec. 3.3.1).
  • [Experiments and practical impact] The evaluation is broad: 7 long-term benchmarks, M4 short-term, few-shot (10%/5%), zero-shot ETT transfer and noise robustness (Tables 1–5, Fig. 4).
  • [Experiments and practical impact] Removing metadata gives the largest drop (ETTh2 MSE 0.272 → 0.324, Table 3), which supports the metadata branch.
  • [Theoretical and application-level implications] The results suggest that structural images and instance-level metadata carry information beyond the raw series, which is a useful signal for multimodal forecasting research.
Critical weaknesses
  • [Importance of the research problem] The "evidence-theory-driven" framing reads more like a label than a motivation drawn from the problem. The paper does not say which uncertainties or conflicts in multimodal forecasting need evidence theory to resolve.
  • [Novelty of the approach] Dempster–Shafer methods for fusion and uncertainty modeling are well established, including in time series work, so the novelty rests mainly on the combination.
  • [Novelty of the approach] There is no comparison with multimodal or vision-based forecasters (Time-VLM, which the paper cites; VisionTS), so the added value over existing multimodal methods cannot be judged.
  • [Soundness and feasibility] Removing the evidential module barely changes results (ETTh1 MSE 0.365 vs. 0.373, Table 3), so the emphasized core module may not drive performance.
  • [Soundness and feasibility] The "BPAs" are unsupervised softmax outputs over 7 focal elements (Sec. 3.2.1), with no Dempster combination rule. In practice this is learnable softmax weighting, not evidence theory in the strict sense.
  • [Soundness and feasibility] The adversarial, sparsity and orthogonality losses (Eq. 2) are not ablated individually.
  • [Experiments and practical impact] The baselines are mainly unimodal numerical models, and comparable multimodal methods are missing.
  • [Experiments and practical impact] EviTime uses a 512-step input window (Sec. 4.4), but the baselines' input lengths are not stated, so the comparison may be unfair.
  • [Experiments and practical impact] There are no means or variances over random seeds, even though some margins are small.
  • [Theoretical and application-level implications] The evidence-theoretic claims are not justified: nothing shows that mass on the "unreliable" state tracks real corrupted observations.
  • [Theoretical and application-level implications] Deployment cost goes unquantified. The method uses a frozen, billion-parameter-scale BLIP-2 (App. B.3) and builds a GAF image per variable at O(B·C·L²) cost (Eq. 22), and the conclusion itself names this cost, but no runtime or memory is reported. Averaging images over hundreds of channels may also wash out structure.
Questions for authors
  1. What input length did each baseline use in Tables 1, 2, 4 and 5, and were they rerun under the same protocol? Can you add Time-VLM, VisionTS, Time-LLM/GPT4TS and iTransformer at L = 512?
  2. Is there any evidence that the learned BPAs reflect observation reliability, e.g., higher mass on {e3} or composite sets for injected corruptions?
  3. Is BLIP-2 fully frozen (only Q-Former/router trained)? What is the trainable parameter count and wall-clock cost per epoch compared with PatchTST?
  4. How does performance change if GVFE is replaced by the same pretrained ViT on a single-channel line plot, or if BLIP-2 is replaced by a random-initialized encoder of equal size? This would separate pretraining from added capacity.
  5. Please report means ± std over at least 3 seeds for Table 1.
  6. In the zero-shot setting (Table 5), dataset-level metadata (domain, frequency) differs between source and target. Is the target's metadata used at test time, and how much does it contribute?
8 comments
AuthorsOpenReview ↗

Thank you for the detailed review. We clarify the evidential formulation, methodological novelty, TEFE contributions, baseline fairness, and zero-shot metadata usage.

W1/Q2. Evidence theory is introduced to characterize observation reliability, ambiguity, and conflicting information in time-series forecasting. Real-world observations may contain noise, local perturbations, and inconsistent temporal or variable-level patterns. EviTime addresses these uncertainties by learning BPAs over seven singleton and composite focal elements from temporal and variable perspectives. Composite focal elements represent ambiguous evidence, while evidence-consistency, strength terms, and Deng entropy support uncertainty-aware fusion. Since BPAs are learned without explicit reliability supervision, their masses represent evidential support rather than calibrated corruption probabilities. The original noise-robustness experiments provide indirect evidence of the practical value of this uncertainty-aware representation.

W2. EviTime's novelty lies in the coordinated modeling of evidential uncertainty, visual temporal structures, and hierarchical contextual semantics rather than DST itself. TEFE constructs uncertainty-aware representations through temporal- and variable-level evidence modeling. GVFE captures complementary temporal patterns using three-channel GAF representations, while MLFE incorporates dataset-, task-, and sample-level information. Cross-attention and gating integrate these complementary representations for forecasting. The original component ablations and additional MLP controls in Table R3-1 support the contributions of the proposed modeling mechanisms beyond conventional statistical features and model capacity.

W3. The soundness of TEFE is supported by its distinct modeling objective, evidential formulation, and fine-grained ablations. Although removing TEFE yields relatively modest performance degradation, it consistently increases forecasting errors, indicating a measurable contribution to uncertainty-aware forecasting. TEFE adopts a Dempster–Shafer-inspired formulation rather than the complete classical DST framework. Its learned BPAs assign mass to singleton and composite focal elements, while Deng entropy characterizes evidential uncertainty without requiring Dempster's combination rule. To further examine the auxiliary objectives in Eq. (2), we individually ablate the adversarial, sparsity, and orthogonality losses, as shown in Table R3-1 of the next comment.

W4/Q1/Q5. We clarify baseline comparability and strengthen experimental reliability through additional forecasting comparisons and random-seed evaluations. EviTime uses a look-back window of 512, while baseline input lengths follow their respective experimental configurations. We acknowledge that input-length differences should be considered when interpreting comparative results. To extend the coverage of recent forecasting methods, we additionally evaluate iTransformer, FreqCycle, and Time-VLM across seven datasets and four forecasting horizons, as shown in Table R3-2. We also repeat the experiments using three random seeds (2024, 2025, and 2026) to assess reproducibility, with results reported in Table R3-3.

W5/Q3. We clarify the evidential interpretation, computational complexity, and practical feasibility of EviTime. TEFE learns uncertainty-aware representations rather than explicitly detecting corrupted observations. Therefore, unreliable-state mass should not be interpreted as a calibrated corruption probability. Although GAF construction has quadratic complexity and variable averaging may suppress certain variable-specific patterns, TEFE retains complementary variable-level information, while aggregation avoids separate visual encoding for each variable. BLIP-2 serves as a frozen pretrained backbone, with 46,351,849 parameters outside the backbone. Since PatchTST differs substantially in architecture and parameter scale, we compare EviTime with Time-VLM, which shares the same BLIP-2 backbone and enables a more controlled assessment of multimodal computational overhead, as reported in Table R3-4.

Q4. Pretrained visual representations contribute to forecasting beyond encoder architecture and parameter capacity. We replace the pretrained BLIP-2 visual encoder with a randomly initialized counterpart of identical architecture and parameter scale. The controlled comparison is presented in Table R3-5.

Q6. Hierarchical metadata provides contextual information without introducing target leakage in zero-shot forecasting. EviTime incorporates dataset-, task-, and sample-level metadata to characterize domain properties, forecasting configurations, and instance-specific statistics. Sample-level descriptors are derived exclusively from observed histories, without accessing future observations or target labels. Since zero-shot forecasting permits observable target-domain characteristics without target-domain training, such metadata is consistent with the zero-shot evaluation protocol. The original metadata ablation further supports its contribution to forecasting performance.

AuthorsOpenReview ↗

Thank you for the constructive review. We conduct auxiliary-loss ablations and MLP controls, and expand forecasting comparisons across seven datasets and four horizons.

Table R3-1. Fine-Grained Loss Ablations and MLP Controls

We ablate three auxiliary losses and compare two MLP controls on four ETT datasets (H = 96). Results are reported as MSE/MAE, with the best values highlighted in bold.

Experimental SettingETTh1ETTh2ETTm1ETTm2
MetricMSE/MAEMSE/MAEMSE/MAEMSE/MAE
Full EviTime0.365/0.3910.272/0.3370.298/0.3430.163/0.253
w/o Adversarial Loss0.367/0.3930.275/0.3390.301/0.3450.165/0.255
w/o Sparsity Loss0.368/0.3930.277/0.3400.302/0.3460.165/0.255
w/o Orthogonality Loss0.367/0.3920.275/0.3390.300/0.3450.164/0.254
Capacity-Matched Gated MLP0.369/0.3950.279/0.3410.303/0.3480.166/0.256
Statistical-Feature MLP0.376/0.4010.293/0.3490.314/0.3560.169/0.259

Analysis: Removing any auxiliary loss increases MSE and MAE across all four datasets, with MSE increases of 0.001–0.005. Removing sparsity loss produces the largest MSE degradation, reaching 0.005 on ETTh2. Compared with the capacity-matched and statistical-feature MLP controls, EviTime reduces MSE by 1.08%–2.51% and 3.55%–7.17%, respectively, supporting the effectiveness of its representation mechanisms.

Table R3-2. Long-term Forecasting Comparison with Recent Baselines

We compare EviTime with three recent baselines across seven datasets and four forecasting horizons (96–720). Results are reported as MSE/MAE, with the best values highlighted in bold.

DatasetHorizonEviTimeiTransformerFreqCycleTime-VLM
MetricMSE/MAEMSE/MAEMSE/MAEMSE/MAE
ETTh1960.365/0.3910.386/0.4050.369/0.3900.361/0.386
1920.406/0.4190.441/0.4360.417/0.4180.397/0.415
3360.434/0.4410.487/0.4580.459/0.4390.420/0.421
7200.449/0.4670.503/0.4910.456/0.4540.441/0.458
Avg0.414/0.4300.454/0.4470.425/0.4250.405/0.420
ETTh2960.272/0.3370.297/0.3490.282/0.3360.267/0.335
1920.345/0.3890.380/0.4000.361/0.3880.326/0.373
3360.360/0.4050.428/0.4320.415/0.4280.357/0.406
7200.398/0.4360.427/0.4450.425/0.4430.412/0.449
Avg0.344/0.3920.383/0.4070.371/0.3990.341/0.391
ETTm1960.298/0.3430.334/0.3680.310/0.3530.304/0.346
1920.334/0.3630.377/0.3910.351/0.3740.332/0.366
3360.363/0.3820.426/0.4200.383/0.3960.364/0.383
7200.428/0.4190.491/0.4590.446/0.4340.402/0.410
Avg0.356/0.3770.407/0.4100.372/0.3890.350/0.377
ETTm2960.163/0.2530.180/0.2640.162/0.2440.160/0.250
1920.222/0.2950.250/0.3090.226/0.2860.215/0.291
3360.273/0.3270.311/0.3480.283/0.3260.270/0.325
7200.357/0.3770.412/0.4070.383/0.3870.348/0.378
Avg0.254/0.3130.288/0.3320.263/0.3110.248/0.311
Weather960.149/0.1920.174/0.2140.159/0.2030.148/0.200
1920.194/0.2360.221/0.2540.208/0.2480.193/0.240
3360.240/0.2690.278/0.2960.264/0.2890.243/0.281
7200.318/0.3270.358/0.3470.343/0.3420.312/0.332
Avg0.225/0.2560.258/0.2780.243/0.2700.224/0.263
Electricity960.137/0.2320.148/0.2400.140/0.2340.142/0.245
1920.149/0.2400.162/0.2530.154/0.2460.157/0.260
3360.165/0.2590.178/0.2690.170/0.2640.174/0.276
7200.193/0.2850.225/0.3170.210/0.2970.214/0.308
Avg0.161/0.2540.178/0.2700.168/0.2600.172/0.273
Traffic960.394/0.2700.395/0.2680.438/0.2620.393/0.290
1920.401/0.2810.417/0.2760.434/0.2580.405/0.296
3360.411/0.2770.433/0.2830.446/0.2580.420/0.305
7200.445/0.2940.467/0.3020.476/0.2650.459/0.323
Avg0.413/0.2810.428/0.2820.448/0.2610.419/0.303

Analysis: Across 28 forecasting settings, EviTime achieves lower MSE than iTransformer and FreqCycle in 28 and 27 cases, respectively. Compared with Time-VLM, it achieves lower MSE in 11 cases and lower MAE in 18 cases. EviTime also achieves the best average MSE/MAE on Electricity (0.161/0.254), average MAE on Weather (0.256), and average MSE on Traffic (0.413), demonstrating competitive performance across forecasting scenarios.

AuthorsOpenReview ↗

Thank you for the valuable suggestions. We conduct three-seed stability tests, computational comparisons with Time-VLM, and pretrained visual encoder controls.

Table R3-3. Stability Analysis Across Three Random Seeds

We evaluate EviTime using three random seeds (2024, 2025, and 2026) across seven datasets. Results are reported as horizon-averaged MSE/MAE for each seed, together with the mean ± sample standard deviation across three seeds.

DatasetSeed 2024Seed 2025Seed 2026Mean ± Std
MetricMSE/MAEMSE/MAEMSE/MAEMSE/MAE
ETTh10.414/0.4300.420/0.4350.409/0.4280.4143±0.0055 / 0.4310±0.0036
ETTh20.344/0.3920.350/0.3970.341/0.3900.3450±0.0046 / 0.3930±0.0036
ETTm10.356/0.3770.361/0.3810.354/0.3750.3570±0.0036 / 0.3777±0.0031
ETTm20.254/0.3130.258/0.3160.252/0.3120.2547±0.0031 / 0.3137±0.0021
Weather0.225/0.2560.230/0.2600.223/0.2540.2260±0.0036 / 0.2567±0.0031
Electricity0.161/0.2540.164/0.2580.160/0.2530.1617±0.0021 / 0.2550±0.0026
Traffic0.413/0.2810.421/0.2860.409/0.2790.4143±0.0061 / 0.2820±0.0036

Analysis: Across seven datasets and three random seeds, MSE standard deviations range from 0.0021 to 0.0061, while MAE standard deviations range from 0.0021 to 0.0036. The largest MSE variation occurs on Traffic (0.0061), whereas Electricity exhibits the smallest MSE variation (0.0021). These consistently low deviations indicate stable forecasting performance across the evaluated random initializations.

Table R3-4. Parameter Scale and Computational Comparison with Time-VLM

We compare EviTime and Time-VLM in terms of parameter scale, inference latency, and peak GPU memory to assess the computational overhead of multimodal modeling.

ModelTotal ParametersBLIP-2 ParametersInference Time (ms/batch)Peak Memory (MiB)
EviTime3,791,113,7053,744,761,856613.31619,243.118
Time-VLM3,745,012,8703,744,761,856595.36919,502.126

Analysis: Compared with Time-VLM, EviTime introduces approximately 46.10M additional parameters, corresponding to a 1.23% increase in total model size. Its inference time increases by 3.01% (613.316 vs. 595.369 ms/batch), while peak GPU memory decreases by 1.33% (19,243.118 vs. 19,502.126 MiB). These results indicate relatively modest additional computational overhead under the same BLIP-2 backbone.

Table R3-5. Effect of Visual Encoder Pretraining

We compare EviTime with a variant that replaces the pretrained BLIP-2 visual encoder with a randomly initialized encoder of the same architecture and size. Results are reported as MSE/MAE (H = 96).

DatasetEviTime (Pretrained)EviTime (Random Init.)
MetricMSE/MAEMSE/MAE
ETTh10.365/0.3910.402/0.419
ETTh20.272/0.3370.331/0.373
ETTm10.298/0.3430.347/0.381
ETTm20.163/0.2530.185/0.273

Analysis: Replacing the pretrained visual encoder with a randomly initialized counterpart increases MSE by 10.14%, 21.69%, 16.44%, and 13.50% on ETTh1, ETTh2, ETTm1, and ETTm2, respectively. MAE also increases consistently across all four datasets, with the largest degradation observed on ETTh2. These results confirm the contribution of pretrained visual representations under identical encoder architectures and parameter scales.

Reviewer_mGq6OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EviTime, a multimodal time series forecasting framework that combines DST-based evidential reasoning with visual and textual modalities. It first generates basic probability assignments from temporal and variable perspectives, fuses the two evidential views using Deng entropy and a learnable gate, and extracts uncertainty-aware temporal features. It also transforms time series into three-channel GAF images for visual encoding via a pretrained BLIP-2 backbone and constructs hierarchical metadata prompts for language encoding. The temporal, visual, and textual representations are aligned and fused for forecasting. Experiments evaluate long-term forecasting on seven benchmarks, noise robustness, few-shot learning, and zero-shot cross-dataset transfer, with ablations on DST, GAF, and metadata.

Critical strengths
  1. Effective cross-granularity zero-shot transferability: In Table 9, EviTime shows consistent zero-shot transfer performance across different ETT subsets (e.g., from hourly ETTh1 to 15-minute ETTm2), indicating that combining structural 2D representations with metadata descriptions can facilitate knowledge transfer across different sampling granularities without target-domain fine-tuning.

  2. Extensive and broad empirical evaluation suite: The paper evaluates the model across multiple standard forecasting setups, including long-term forecasting over various horizons, short-term forecasting on M4, input noise perturbations, few-shot forecasting, and zero-shot cross-dataset transfer. The supplementary materials provide exhaustive horizon-by-horizon numerical tables, demonstrating a commendable experimental effort.

Critical weaknesses
  1. Several components of EviTime have precedents: DST for time series (e.g., TEFN, Zhan et al., 2025), GAF for time-series-to-image conversion (widely used since 2015), and statistical prompts for LLMs (e.g., Time-LLM). The paper does not provide a deep analysis of the synergistic mechanism among DST, GAF, and metadata. Without such justification, the combination may appear as a simple concatenation of existing modules rather than a principled multimodal framework.

  2. EviTime uses BLIP-2, a billion-parameter vision-language model, but mainly compares against lightweight numerical baselines such as DLinear, PatchTST, TimesNet, and Autoformer. It cites multimodal/foundation-model baselines like Time-VLM and M3Time but does not compare against them. Thus, the reported gains may stem from the pretrained VLM’s capacity rather than the proposed evidential, GAF, or metadata designs. Same-capacity baselines and parameter/compute comparisons are needed.

  3. Under 5% training data on high-dimensional datasets, EviTime is outperformed by simpler or lighter-weight baselines: on Electricity, the linear model DLinear achieves 0.176 MSE vs. EviTime’s 0.194; on Traffic, the transformer-based PatchTST achieves 0.418 vs. EviTime’s 0.470.Moreover,removing DST only slightly hurts performance (ETTh1: 0.365→0.373; ETTm2: 0.163→0.169), while removing metadata causes the largest drop. This questions whether the DST module is essential and whether the architecture’s complexity is justified.

Questions for authors
  1. Several components of EviTime have precedents: DST for time series (e.g., TEFN, Zhan et al., 2025), GAF for time-series-to-image conversion (widely used since 2015), and statistical prompts for LLMs (e.g., Time-LLM). Could the authors clarify what is fundamentally new beyond combining these established techniques, and explain the deeper synergistic mechanism among DST, GAF, and metadata that justifies their joint use?

  2. Baseline Fairness: Can the authors provide direct comparisons under identical benchmark protocols with at least one representative multimodal or foundation-model baseline (e.g., Time-VLM or UniTS) to isolate the contribution of the proposed framework from the raw parameter scale of BLIP-2?

  3. Under 5% training data on Electricity and Traffic, EviTime is outperformed by DLinear (0.176 vs. 0.194 MSE) and PatchTST (0.418 vs. 0.470 MSE). How do you explain this, and in what data regimes does EviTime’s advantage hold?

7 comments
AuthorsOpenReview ↗

Thank you for the constructive comments and valuable suggestions. We further clarify EviTime's methodological novelty, baseline fairness, and performance under limited training data.

Specifically, we address three aspects: (1) the synergistic mechanism among DST, GAF, and hierarchical metadata; (2) the influence of pretrained model capacity and comparisons with representative multimodal forecasting methods; and (3) EviTime's performance under different data regimes and the independent contribution of DST.

W1/Q1. EviTime's fundamental novelty lies in evidence-aware multimodal representation learning, rather than simply combining existing DST, GAF, and metadata techniques. Unlike TEFN, EviTime constructs nonlinear BPAs from both temporal and variable perspectives, using composite focal elements to preserve ambiguous evidence and Deng entropy to characterize uncertainty for adaptive evidential fusion. This produces uncertainty-aware numerical representations instead of relying solely on deterministic temporal features. More importantly, the three modalities address different aspects of the same forecasting problem: DST characterizes evidential support and ambiguity in observations, three-channel GAF captures structural temporal patterns, and hierarchical metadata provides contextual semantics across dataset, task, and sample levels. Their synergy arises from these complementary information sources: evidential reasoning distinguishes uncertain numerical patterns, visual representations capture temporal structures beyond individual observations, and metadata supplies contextual information for interpreting these patterns across forecasting settings. The resulting representations are aligned and integrated for prediction, rather than treated as interchangeable features. The original Table 3 supports their complementary contributions, as removing any component degrades forecasting performance.

W2/Q2. EviTime's effectiveness is not solely attributable to the large pretrained BLIP-2 backbone, as supported by direct multimodal comparisons and computational evaluations. First, BLIP-2 serves as a pretrained representation backbone, while EviTime's DST-based evidential reasoning, three-channel GAF construction, hierarchical metadata, and multimodal fusion introduce distinct modeling mechanisms. The original component ablations demonstrate their contributions within the same framework, rather than attributing improvements solely to pretrained model capacity. Second, to address baseline fairness, we compare with iTransformer, FreqCycle, and the multimodal Time-VLM across seven datasets and four forecasting horizons. The results demonstrate EviTime's competitive performance against recent forecasting methods, including a representative multimodal model. Third, to quantify computational costs, we compare EviTime with Time-VLM using the same 3.745B-parameter BLIP-2 backbone. EviTime has approximately 1.23% more total parameters and requires 3.01% more inference time, but uses 1.33% less peak GPU memory. These results indicate that EviTime achieves competitive forecasting performance with modest additional computational overhead under a shared pretrained backbone. The forecasting comparisons and computational evaluations are presented in Tables R2-1 and R2-2 of the second comment, respectively.

W3/Q3. EviTime's advantages vary across data regimes, while DST serves a distinct purpose in uncertainty-aware forecasting. Under the 5% training setting, DLinear and PatchTST achieve lower MSE on Electricity and Traffic, respectively, indicating that lightweight numerical models can be advantageous when training data are extremely limited. Nevertheless, EviTime demonstrates competitive performance across broader forecasting settings, including long-term prediction and zero-shot cross-dataset transfer. Regarding DST, its modeling purpose is to characterize latent evidential support and ambiguity, rather than merely increase model capacity. Unlike metadata, which provides contextual semantics, DST constructs BPAs from temporal and variable perspectives and employs Deng entropy to characterize uncertainty for adaptive evidence fusion. Thus, their contributions should not be judged solely by relative ablation magnitudes. The original ablation shows that removing DST increases MSE from 0.365 to 0.373 on ETTh1 and from 0.163 to 0.169 on ETTm2, corresponding to increases of 2.19% and 3.68%, respectively. To further examine the contribution of evidential representation learning, we conduct fine-grained ablations by separately removing the adversarial, sparsity, and orthogonality losses, which impose complementary constraints on representation learning. Their removal consistently degrades forecasting performance across four ETT datasets. The fine-grained auxiliary-loss ablation results are presented in Table R2-3 of the second comment.

AuthorsOpenReview ↗

Thank you for the valuable suggestions regarding experimental validation. We present additional baseline comparisons, computational evaluations, and fine-grained ablation experiments.

Table R2-1. Long-term Forecasting Comparison with Recent Baselines

We compare EviTime with three baselines across seven datasets and four horizons. Results are reported as MSE/MAE.

DatasetHorizonEviTimeiTransformerFreqCycleTime-VLM
MetricMSE/MAEMSE/MAEMSE/MAEMSE/MAE
ETTh1960.365/0.3910.386/0.4050.369/0.3900.361/0.386
1920.406/0.4190.441/0.4360.417/0.4180.397/0.415
3360.434/0.4410.487/0.4580.459/0.4390.420/0.421
7200.449/0.4670.503/0.4910.456/0.4540.441/0.458
Avg0.414/0.4300.454/0.4470.425/0.4250.405/0.420
ETTh2960.272/0.3370.297/0.3490.282/0.3360.267/0.335
1920.345/0.3890.380/0.4000.361/0.3880.326/0.373
3360.360/0.4050.428/0.4320.415/0.4280.357/0.406
7200.398/0.4360.427/0.4450.425/0.4430.412/0.449
Avg0.344/0.3920.383/0.4070.371/0.3990.341/0.391
ETTm1960.298/0.3430.334/0.3680.310/0.3530.304/0.346
1920.334/0.3630.377/0.3910.351/0.3740.332/0.366
3360.363/0.3820.426/0.4200.383/0.3960.364/0.383
7200.428/0.4190.491/0.4590.446/0.4340.402/0.410
Avg0.356/0.3770.407/0.4100.372/0.3890.350/0.377
ETTm2960.163/0.2530.180/0.2640.162/0.2440.160/0.250
1920.222/0.2950.250/0.3090.226/0.2860.215/0.291
3360.273/0.3270.311/0.3480.283/0.3260.270/0.325
7200.357/0.3770.412/0.4070.383/0.3870.348/0.378
Avg0.254/0.3130.288/0.3320.263/0.3110.248/0.311
Weather960.149/0.1920.174/0.2140.159/0.2030.148/0.200
1920.194/0.2360.221/0.2540.208/0.2480.193/0.240
3360.240/0.2690.278/0.2960.264/0.2890.243/0.281
7200.318/0.3270.358/0.3470.343/0.3420.312/0.332
Avg0.225/0.2560.258/0.2780.243/0.2700.224/0.263
Electricity960.137/0.2320.148/0.2400.140/0.2340.142/0.245
1920.149/0.2400.162/0.2530.154/0.2460.157/0.260
3360.165/0.2590.178/0.2690.170/0.2640.174/0.276
7200.193/0.2850.225/0.3170.210/0.2970.214/0.308
Avg0.161/0.2540.178/0.2700.168/0.2600.172/0.273
Traffic960.394/0.2700.395/0.2680.438/0.2620.393/0.290
1920.401/0.2810.417/0.2760.434/0.2580.405/0.296
3360.411/0.2770.433/0.2830.446/0.2580.420/0.305
7200.445/0.2940.467/0.3020.476/0.2650.459/0.323
Avg0.413/0.2810.428/0.2820.448/0.2610.419/0.303

Analysis: Across 28 forecasting settings, EviTime achieves lower MSE than iTransformer and FreqCycle in 28 and 27 cases, respectively. Compared with Time-VLM, it achieves lower MSE in 11 cases and lower MAE in 18 cases. EviTime also achieves the best average MSE/MAE on Electricity, average MAE on Weather, and average MSE on Traffic.

Table R2-2. Parameter Scale and Computational Comparison with Time-VLM

We compare total parameters, BLIP-2 backbone size, inference time, and peak GPU memory.

ModelTotal ParametersBLIP-2 ParametersInference Time (ms/batch)Peak Memory (MiB)
EviTime3,791,113,7053,744,761,856613.31619,243.118
Time-VLM3,745,012,8703,744,761,856595.36919,502.126

Analysis: Compared with Time-VLM, EviTime has 46.10M more total parameters and 3.01% higher inference latency, while its peak GPU memory is 1.33% lower. This comparison quantifies the accuracy–efficiency trade-off under a shared pretrained backbone.

Table R2-3. Fine-Grained Ablation Study of Auxiliary Losses

We ablate three losses on four ETT datasets. Results are reported as MSE/MAE.

Experimental SettingETTh1ETTh2ETTm1ETTm2
MetricMSE/MAEMSE/MAEMSE/MAEMSE/MAE
Full EviTime0.365/0.3910.272/0.3370.298/0.3430.163/0.253
w/o Adversarial Loss0.367/0.3930.275/0.3390.301/0.3450.165/0.255
w/o Sparsity Loss0.368/0.3930.277/0.3400.302/0.3460.165/0.255
w/o Orthogonality Loss0.367/0.3920.275/0.3390.300/0.3450.164/0.254

Analysis: Removing any auxiliary loss increases both MSE and MAE. MSE increases range from 0.001 to 0.005, with sparsity-loss removal producing the largest increase of 0.005 on ETTh2.

AuthorsOpenReview ↗

A Brief Summary of Our Earlier Responses on Methodological Novelty, Component Contributions, and Experimental Validation

We sincerely thank the reviewer for the constructive comments regarding the methodological novelty, component contributions, and experimental validation of EviTime. We would like to briefly summarize the clarifications and additional experimental evidence presented in our earlier responses.

First, we clarified that EviTime aims to achieve uncertainty-aware multimodal forecasting through evidential representation learning, rather than simply combining existing techniques. Specifically, the DST-inspired branch models evidential support and uncertainty through temporal and variable perspectives, while composite focal elements preserve ambiguity and Deng entropy guides adaptive evidence fusion.

Second, we further explained the complementary roles of the proposed components. The three-channel GAF representation captures temporal similarity, directional variations, and amplitude dynamics, while hierarchical metadata provides dataset-, task-, and sample-level semantic information. Together, these components address distinct modeling challenges and contribute to more comprehensive multimodal representations.

Third, we extended the experimental comparisons by incorporating three recent baselines, including iTransformer, FreqCycle, and Time-VLM, across seven datasets and 28 forecasting settings. The additional results demonstrate the competitive performance of EviTime across different forecasting scenarios, further strengthening the experimental validation.

Fourth, we conducted fine-grained ablations of three auxiliary losses and introduced capacity-matched gated MLP and statistical-feature MLP controls. The results consistently support the contributions of the auxiliary objectives and demonstrate that EviTime's improvements cannot be attributed solely to increased model capacity or conventional statistical feature encoding. We also clarified the distinct purposes of component ablation and noise robustness evaluation.

We hope these additional explanations and experiments have adequately addressed the reviewer's concerns and further clarified the methodological contributions and effectiveness of EviTime. We would greatly appreciate any further feedback from the reviewer at their earliest convenience, as it would help us promptly address any remaining questions.

We sincerely thank the reviewer and the Area Chair for their time, thoughtful consideration, and continued evaluation of our work.

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

This paper proposes EviTime, a multimodal time series forecasting framework that combines Dempster-Shafer Theory (DST)-based evidential representations, three-channel GAF visual features, and multi-granular metadata text prompts via a pretrained BLIP-2 backbone. The extracted numerical, visual, and semantic features are fused to perform long-term, short-term, few-shot, and zero-shot forecasting.

Critical strengths
  1. Comprehensive experimental scope covering long-term, short-term, few-shot, zero-shot transfer, and noise robustness evaluations.

  2. Utilizing hierarchical metadata (domain descriptions, task settings, and sample-level statistics) is a practical and reusable way to inject contextual priors into forecasting.

Critical weaknesses
  1. Inadequate and outdated baseline selection. Most compared baselines date back to 2021–2022 (e.g., Autoformer, FEDformer, ETSformer, LightTS). The paper omits classic strong baselines for multivariate correlation modeling such as iTransformer, as well as recent state-of-the-art forecasting models from 2026 (e.g., FreqCycle [AAAI 2026] and Time-Align[ICLR 2026]). More importantly, despite citing closely related multimodal/VLM forecasting works such as Time-VLM, T3Time, and M3Time, none of them are included in the experiments, making the claimed empirical superiority unconvincing.

  2. Coarse-grained ablations fail to justify the stacked sub-modules in both DST and VLM branches. The paper only conducts macro-level ablations by removing entire branches (w/o DST, w/o GAF, w/o Metadata), leaving numerous internal tricks unverified. Specifically, in the DST branch (TEFE), the authors stack dual-view BPA generation, evidence consistency/strength weighting, Deng-entropy gating, and three auxiliary losses (adversarial, sparsity, and orthogonality), yet provide no fine-grained ablations to show these complex designs outperform a simple gated MLP or that each sub-component is necessary. Similarly, in the VLM branches (GVFE & MLFE), there are no ablations on individual GAF channels (GASF vs. GADF vs. amplitude) or metadata router tokens; the massive drop in w/o Metadata also fails to separate whether the gain comes from BLIP-2's semantic understanding or simply from feeding raw summary statistics (mean, std, min, max). Moreover, the noise robustness test lacks any baseline or w/o DST comparison.

  3. Limited novelty and unverified evidential semantics. Since TEFN already models dual-view BPAs for time series and Time-VLM already integrates vision-language backbones, this work reads largely as an ensemble of known paradigms. Moreover, the DST module lacks structural constraints or Dempster's combination rule to guarantee that the Softmax outputs actually learn the claimed reliability states rather than acting as a generic gated MLP, and minimizing Deng entropy directly contradicts the goal of modeling multi-state ambiguity.

Questions for authors
  1. Given the outdated baseline pool, I doubt that the claimed SOTA margins hold against stronger contemporary methods. To truly validate the empirical advantage, the authors should compare EviTime with iTransformer (especially on high-dimensional benchmarks like Traffic where iTransformer is known to perform strongly), recent 2026 models (e.g., FreqCycle, TimeAlign), and direct multimodal competitors (e.g., Time-VLM)
  2. To solidify the architectural design, fine-grained ablations are needed beyond macro-branch removal. Specifically, the authors should evaluate the individual contributions of the adversarial, sparsity, and orthogonality losses, replace the DST module with a standard gated MLP of equal capacity, and compare the VLM text branch against feeding the sample-level statistical features directly into a lightweight MLP.
  3. Currently, the DST module reads like mathematical packaging without empirical verification of its internal behavior. To substantiate the evidential claims, the authors could provide supportive interpretability analyses—such as visualizing the learned BPA mass distributions, Deng entropy, and gating weights across clean vs. noisy/corrupted intervals—to demonstrate that the module genuinely captures observation reliability rather than collapsing into a degenerate one-hot assignment.
7 comments
AuthorsOpenReview ↗

Thank you for the constructive comments. We clarify the roles of evidential uncertainty, GAF representations, and hierarchical metadata.

We address three aspects: (1) the effectiveness of EviTime across different forecasting paradigms and evaluation settings, (2) the complementary roles of DST, GAF, hierarchical metadata, and auxiliary objectives, together with the distinct purposes of component ablation and noise robustness evaluation, and (3) the methodological differences from existing approaches, particularly the interpretation of BPAs, composite focal elements, and Deng entropy in uncertainty-aware representation learning.

W1 & Q1. EviTime has already been evaluated against recent and representative methods from different forecasting paradigms, supporting the effectiveness of its evidence-aware multimodal framework. Unlike conventional forecasting models that primarily learn temporal dependencies, EviTime integrates Dempster–Shafer-inspired evidential modeling with visual and textual representations to characterize uncertainty and enhance multimodal feature integration. Our existing experiments include comparisons with representative methods across different model categories and forecasting settings, covering long-term, short-term, few-shot, and zero-shot forecasting. These evaluations provide evidence of EviTime's effectiveness across the evaluated scenarios. Incorporating additional recent baselines further strengthens the comprehensiveness of the evaluation. Therefore, following the reviewer's suggestion, we have extended our comparisons to include iTransformer, FreqCycle, and Time-VLM, with the additional experimental results reported in Table R1-1 of the second comment.

W2/Q2. EviTime's components provide complementary representations rather than merely increasing model complexity. The DST-inspired temporal branch models evidential support and uncertainty through BPAs constructed from temporal and variable perspectives, while the three-channel GAF representation captures temporal similarity, directional variations, and amplitude dynamics. Hierarchical metadata introduces dataset-, task-, and sample-level semantics through instance-relevant textual information. These components address distinct modeling challenges, and the original Table 3 demonstrates their complementary contributions through component-level ablations. Moreover, the auxiliary objectives impose different constraints on representation learning rather than simply increasing model capacity. To further improve the interpretability of these contributions, we have conducted fine-grained loss ablations and controlled comparisons with a capacity-matched gated MLP and a statistical-feature MLP, with results reported in Table R1-2 of the second comment. Regarding noise robustness, Figure 4 evaluates EviTime under six progressively increasing noise levels across four ETT datasets to characterize its sensitivity to input perturbations, rather than isolate individual components. This evaluation complements the DST ablation and controlled comparisons, which examine architectural contributions and predictive effectiveness. Together, these experiments assess distinct but complementary aspects of model performance without conflating robustness evaluation with component attribution.

W3/Q3. EviTime's novelty lies in modeling evidential uncertainty within multimodal forecasting rather than simply combining existing techniques. Unlike TEFN, which focuses on temporal feature modeling, and Time-VLM, which leverages visual and textual representations, EviTime constructs nonlinear BPAs from temporal and variable perspectives to model evidential support and uncertainty. Its three hypotheses represent reliable, locally perturbed, and unreliable latent states, while composite focal elements preserve ambiguity when evidence cannot be confidently assigned to a single hypothesis. These learned assignments characterize latent evidential states rather than calibrated observation reliability or corruption probabilities. Deng entropy quantifies uncertainty across these belief assignments and guides adaptive evidence fusion, enabling uncertainty-aware representations rather than deterministic feature integration. Importantly, the original DST ablation demonstrates the predictive effectiveness of evidential modeling, while the additional capacity-matched and statistical-feature controls in Table R1-2 of the second comment further support its contribution beyond increased model capacity or conventional feature encoding.

AuthorsOpenReview ↗

Thank you for the valuable suggestions regarding experimental validation. We conduct additional baseline comparisons and fine-grained ablation experiments.

Table R1-1. Long-term Forecasting Comparison with Recent Baselines

We compare EviTime with three recent baselines across seven datasets and four forecasting horizons (96–720). Results are reported as MSE/MAE, with the best values highlighted in bold.

DatasetHorizonEviTimeiTransformerFreqCycleTime-VLM
MetricMSE/MAEMSE/MAEMSE/MAEMSE/MAE
ETTh1960.365/0.3910.386/0.4050.369/0.3900.361/0.386
1920.406/0.4190.441/0.4360.417/0.4180.397/0.415
3360.434/0.4410.487/0.4580.459/0.4390.420/0.421
7200.449/0.4670.503/0.4910.456/0.4540.441/0.458
Avg0.414/0.4300.454/0.4470.425/0.4250.405/0.420
ETTh2960.272/0.3370.297/0.3490.282/0.3360.267/0.335
1920.345/0.3890.380/0.4000.361/0.3880.326/0.373
3360.360/0.4050.428/0.4320.415/0.4280.357/0.406
7200.398/0.4360.427/0.4450.425/0.4430.412/0.449
Avg0.344/0.3920.383/0.4070.371/0.3990.341/0.391
ETTm1960.298/0.3430.334/0.3680.310/0.3530.304/0.346
1920.334/0.3630.377/0.3910.351/0.3740.332/0.366
3360.363/0.3820.426/0.4200.383/0.3960.364/0.383
7200.428/0.4190.491/0.4590.446/0.4340.402/0.410
Avg0.356/0.3770.407/0.4100.372/0.3890.350/0.377
ETTm2960.163/0.2530.180/0.2640.162/0.2440.160/0.250
1920.222/0.2950.250/0.3090.226/0.2860.215/0.291
3360.273/0.3270.311/0.3480.283/0.3260.270/0.325
7200.357/0.3770.412/0.4070.383/0.3870.348/0.378
Avg0.254/0.3130.288/0.3320.263/0.3110.248/0.311
Weather960.149/0.1920.174/0.2140.159/0.2030.148/0.200
1920.194/0.2360.221/0.2540.208/0.2480.193/0.240
3360.240/0.2690.278/0.2960.264/0.2890.243/0.281
7200.318/0.3270.358/0.3470.343/0.3420.312/0.332
Avg0.225/0.2560.258/0.2780.243/0.2700.224/0.263
Electricity960.137/0.2320.148/0.2400.140/0.2340.142/0.245
1920.149/0.2400.162/0.2530.154/0.2460.157/0.260
3360.165/0.2590.178/0.2690.170/0.2640.174/0.276
7200.193/0.2850.225/0.3170.210/0.2970.214/0.308
Avg0.161/0.2540.178/0.2700.168/0.2600.172/0.273
Traffic960.394/0.2700.395/0.2680.438/0.2620.393/0.290
1920.401/0.2810.417/0.2760.434/0.2580.405/0.296
3360.411/0.2770.433/0.2830.446/0.2580.420/0.305
7200.445/0.2940.467/0.3020.476/0.2650.459/0.323
Avg0.413/0.2810.428/0.2820.448/0.2610.419/0.303

Analysis: Across 28 forecasting settings, EviTime achieves lower MSE than iTransformer and FreqCycle in 28 and 27 cases, respectively. Compared with Time-VLM, it achieves lower MSE in 11 cases and lower MAE in 18 cases. EviTime also achieves the best average MSE/MAE on Electricity (0.161/0.254), average MAE on Weather (0.256), and average MSE on Traffic (0.413), demonstrating competitive performance across forecasting scenarios.

Table R1-2. Fine-Grained Loss Ablations and MLP Controls

We ablate three auxiliary losses and compare two MLP controls on four ETT datasets (H = 96). Results are reported as MSE/MAE, with the best values highlighted in bold.

Experimental SettingETTh1ETTh2ETTm1ETTm2
MetricMSE/MAEMSE/MAEMSE/MAEMSE/MAE
Full EviTime0.365/0.3910.272/0.3370.298/0.3430.163/0.253
w/o Adversarial Loss0.367/0.3930.275/0.3390.301/0.3450.165/0.255
w/o Sparsity Loss0.368/0.3930.277/0.3400.302/0.3460.165/0.255
w/o Orthogonality Loss0.367/0.3920.275/0.3390.300/0.3450.164/0.254
Capacity-Matched Gated MLP0.369/0.3950.279/0.3410.303/0.3480.166/0.256
Statistical-Feature MLP0.376/0.4010.293/0.3490.314/0.3560.169/0.259

Analysis: Removing any auxiliary loss increases MSE and MAE across all four datasets, with MSE increases ranging from 0.001 to 0.005. Removing sparsity loss produces the largest MSE degradation, reaching 0.005 on ETTh2. Full EviTime also outperforms both MLP controls on all eight metrics, supporting the benefits of its representation mechanisms beyond model capacity and conventional statistical features.

AuthorsOpenReview ↗

A Brief Summary of Our Earlier Responses on Methodological Novelty, Component Contributions, and Experimental Validation

We sincerely thank the reviewer for the constructive comments regarding the methodological novelty, component contributions, and experimental validation of EviTime. We would like to briefly summarize the clarifications and additional experimental evidence presented in our earlier responses.

First, we clarified that EviTime aims to achieve uncertainty-aware multimodal forecasting through evidential representation learning, rather than simply combining existing techniques. Specifically, the DST-inspired branch models evidential support and uncertainty through temporal and variable perspectives, while composite focal elements preserve ambiguity and Deng entropy guides adaptive evidence fusion.

Second, we further explained the complementary roles of the proposed components. The three-channel GAF representation captures temporal similarity, directional variations, and amplitude dynamics, while hierarchical metadata provides dataset-, task-, and sample-level semantic information. Together, these components address distinct modeling challenges and contribute to more comprehensive multimodal representations.

Third, we extended the experimental comparisons by incorporating three recent baselines, including iTransformer, FreqCycle, and Time-VLM, across seven datasets and 28 forecasting settings. The additional results demonstrate the competitive performance of EviTime across different forecasting scenarios, further strengthening the experimental validation.

Fourth, we conducted fine-grained ablations of three auxiliary losses and introduced capacity-matched gated MLP and statistical-feature MLP controls. The results consistently support the contributions of the auxiliary objectives and demonstrate that EviTime's improvements cannot be attributed solely to increased model capacity or conventional statistical feature encoding. We also clarified the distinct purposes of component ablation and noise robustness evaluation.

We hope these additional explanations and experiments have adequately addressed the reviewer's concerns and further clarified the methodological contributions and effectiveness of EviTime. We would greatly appreciate any further feedback from the reviewer at their earliest convenience, as it would help us promptly address any remaining questions.

We sincerely thank the reviewer and the Area Chair for their time, thoughtful consideration, and continued evaluation of our work.

Comments on the paper

2 comments
AuthorsOpenReview ↗

We sincerely thank the reviewers, Program Chairs, Area Chair, and Senior Program Committee members for their careful evaluation of EviPhys and the opportunity to clarify its methodological contributions, theoretical foundations, and experimental reliability.

We sincerely thank the reviewers, Program Chairs, Area Chair, and Senior Program Committee members for their careful evaluation of EviTime and the opportunity to clarify its methodological novelty, evidential formulation, experimental validity, and practical applicability.

Our work makes three complementary contributions to evidence-aware multimodal time-series forecasting. First, at the evidential representation stage, we develop TEFE to construct nonlinear basic probability assignments (BPAs) from temporal and variable perspectives. Singleton and composite focal elements, together with Deng entropy, characterize evidential support and ambiguity for uncertainty-aware numerical representations. Second, at the visual representation stage, we introduce GVFE to construct three-channel Gramian Angular Field (GAF) representations, capturing temporal structures, directional variations, and amplitude dynamics. Third, at the contextual representation and multimodal fusion stage, we design MLFE to incorporate dataset-, task-, and sample-level metadata. Cross-attention and gating integrate these complementary representations for forecasting.

Our responses address the reviewers' concerns regarding methodological novelty, evidential interpretation, experimental fairness, component effectiveness, computational efficiency, and reproducibility.

  1. Methodological novelty and evidential interpretation. We clarify that EviTime coordinates evidential uncertainty, visual temporal structures, and hierarchical contextual semantics rather than simply combining existing techniques. We distinguish our Dempster–Shafer-inspired formulation from classical DST and explain the roles of learned BPAs, composite focal elements, and Deng entropy. Learned evidential masses characterize latent support and ambiguity rather than calibrated corruption probabilities, ensuring a precise interpretation of uncertainty-aware representations.

  2. Component effectiveness and controlled ablation studies. We explain the complementary roles of TEFE, GVFE, and MLFE and distinguish component ablations from noise-robustness evaluations. We conduct fine-grained ablations of adversarial, sparsity, and orthogonality losses, together with comparisons against a capacity-matched gated MLP and a statistical-feature MLP. Removing any auxiliary loss increases forecasting errors across four ETT datasets, while EviTime outperforms both MLP controls, supporting its representation mechanisms beyond statistical features and model capacity.

  3. Baseline fairness and comprehensive forecasting evaluation. We extend comparisons to iTransformer, FreqCycle, and Time-VLM across seven datasets and four horizons, covering 28 forecasting settings. EviTime achieves lower MSE than iTransformer and FreqCycle in 28 and 27 settings, respectively, and lower MSE/MAE than Time-VLM in 11/18 settings. It also achieves the best average MSE/MAE on Electricity, average MAE on Weather, and average MSE on Traffic. We clarify input-length differences and acknowledge the advantages of lightweight models under extremely limited training data.

  4. Pretrained representation contributions and computational efficiency. We distinguish EviTime's modeling contributions from those of pretrained BLIP-2. Replacing the pretrained visual encoder with an identically sized randomly initialized counterpart increases MSE by 10.14%–21.69% across four ETT datasets. Compared with Time-VLM using the same 3.745B-parameter BLIP-2 backbone, EviTime introduces 46.10M additional parameters and 3.01% higher inference latency, while reducing peak GPU memory by 1.33%, clarifying its accuracy–efficiency trade-off.

  5. Experimental reliability, robustness, and zero-shot validity. We conduct three-seed evaluations across seven datasets, reporting horizon-averaged MSE/MAE and sample standard deviations. MSE standard deviations range from 0.0021 to 0.0061, supporting stability across evaluated random initializations. We clarify that noise-robustness experiments assess sensitivity to perturbations rather than individual component contributions. We also explain that zero-shot metadata relies on observable characteristics and historical statistics without accessing future targets or target-domain training labels.

Together, these clarifications and additional experiments strengthen the methodological transparency, experimental fairness, interpretability, reproducibility, and empirical validation of EviTime. We also acknowledge limitations related to pretrained model capacity, input-length differences, and limited training data, ensuring that our conclusions remain consistent with the experimental evidence.

AuthorsOpenReview ↗

Request for Further Discussion

Dear Area Chair, Program Chairs, and Reviewers,

We sincerely appreciate your time and careful consideration of our submission. We have provided detailed responses and substantial additional experimental results addressing the reviewers' concerns regarding methodological novelty, evidential formulation, component effectiveness, baseline fairness, computational efficiency, and experimental reliability. In particular, we have strengthened the evaluation through controlled ablations, additional baseline comparisons, multi-seed experiments, and analyses of pretrained representations and zero-shot validity. As the discussion deadline is approaching, we sincerely hope the reviewers will have an opportunity to examine our responses and share their updated assessments. We would greatly appreciate the Area Chair's and Program Chairs' assistance in encouraging further discussion, allowing us to clarify any remaining concerns before the discussion concludes. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.