- 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
- 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?
- Is there any evidence that the learned BPAs reflect observation reliability, e.g., higher mass on {e3} or composite sets for injected corruptions?
- 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?
- 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.
- Please report means ± std over at least 3 seeds for Table 1.
- 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
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 Setting | ETTh1 | ETTh2 | ETTm1 | ETTm2 |
|---|---|---|---|---|
| Metric | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE |
| Full EviTime | 0.365/0.391 | 0.272/0.337 | 0.298/0.343 | 0.163/0.253 |
| w/o Adversarial Loss | 0.367/0.393 | 0.275/0.339 | 0.301/0.345 | 0.165/0.255 |
| w/o Sparsity Loss | 0.368/0.393 | 0.277/0.340 | 0.302/0.346 | 0.165/0.255 |
| w/o Orthogonality Loss | 0.367/0.392 | 0.275/0.339 | 0.300/0.345 | 0.164/0.254 |
| Capacity-Matched Gated MLP | 0.369/0.395 | 0.279/0.341 | 0.303/0.348 | 0.166/0.256 |
| Statistical-Feature MLP | 0.376/0.401 | 0.293/0.349 | 0.314/0.356 | 0.169/0.259 |
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.
| Dataset | Horizon | EviTime | iTransformer | FreqCycle | Time-VLM |
|---|---|---|---|---|---|
| Metric | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE | |
| ETTh1 | 96 | 0.365/0.391 | 0.386/0.405 | 0.369/0.390 | 0.361/0.386 |
| 192 | 0.406/0.419 | 0.441/0.436 | 0.417/0.418 | 0.397/0.415 | |
| 336 | 0.434/0.441 | 0.487/0.458 | 0.459/0.439 | 0.420/0.421 | |
| 720 | 0.449/0.467 | 0.503/0.491 | 0.456/0.454 | 0.441/0.458 | |
| Avg | 0.414/0.430 | 0.454/0.447 | 0.425/0.425 | 0.405/0.420 | |
| ETTh2 | 96 | 0.272/0.337 | 0.297/0.349 | 0.282/0.336 | 0.267/0.335 |
| 192 | 0.345/0.389 | 0.380/0.400 | 0.361/0.388 | 0.326/0.373 | |
| 336 | 0.360/0.405 | 0.428/0.432 | 0.415/0.428 | 0.357/0.406 | |
| 720 | 0.398/0.436 | 0.427/0.445 | 0.425/0.443 | 0.412/0.449 | |
| Avg | 0.344/0.392 | 0.383/0.407 | 0.371/0.399 | 0.341/0.391 | |
| ETTm1 | 96 | 0.298/0.343 | 0.334/0.368 | 0.310/0.353 | 0.304/0.346 |
| 192 | 0.334/0.363 | 0.377/0.391 | 0.351/0.374 | 0.332/0.366 | |
| 336 | 0.363/0.382 | 0.426/0.420 | 0.383/0.396 | 0.364/0.383 | |
| 720 | 0.428/0.419 | 0.491/0.459 | 0.446/0.434 | 0.402/0.410 | |
| Avg | 0.356/0.377 | 0.407/0.410 | 0.372/0.389 | 0.350/0.377 | |
| ETTm2 | 96 | 0.163/0.253 | 0.180/0.264 | 0.162/0.244 | 0.160/0.250 |
| 192 | 0.222/0.295 | 0.250/0.309 | 0.226/0.286 | 0.215/0.291 | |
| 336 | 0.273/0.327 | 0.311/0.348 | 0.283/0.326 | 0.270/0.325 | |
| 720 | 0.357/0.377 | 0.412/0.407 | 0.383/0.387 | 0.348/0.378 | |
| Avg | 0.254/0.313 | 0.288/0.332 | 0.263/0.311 | 0.248/0.311 | |
| Weather | 96 | 0.149/0.192 | 0.174/0.214 | 0.159/0.203 | 0.148/0.200 |
| 192 | 0.194/0.236 | 0.221/0.254 | 0.208/0.248 | 0.193/0.240 | |
| 336 | 0.240/0.269 | 0.278/0.296 | 0.264/0.289 | 0.243/0.281 | |
| 720 | 0.318/0.327 | 0.358/0.347 | 0.343/0.342 | 0.312/0.332 | |
| Avg | 0.225/0.256 | 0.258/0.278 | 0.243/0.270 | 0.224/0.263 | |
| Electricity | 96 | 0.137/0.232 | 0.148/0.240 | 0.140/0.234 | 0.142/0.245 |
| 192 | 0.149/0.240 | 0.162/0.253 | 0.154/0.246 | 0.157/0.260 | |
| 336 | 0.165/0.259 | 0.178/0.269 | 0.170/0.264 | 0.174/0.276 | |
| 720 | 0.193/0.285 | 0.225/0.317 | 0.210/0.297 | 0.214/0.308 | |
| Avg | 0.161/0.254 | 0.178/0.270 | 0.168/0.260 | 0.172/0.273 | |
| Traffic | 96 | 0.394/0.270 | 0.395/0.268 | 0.438/0.262 | 0.393/0.290 |
| 192 | 0.401/0.281 | 0.417/0.276 | 0.434/0.258 | 0.405/0.296 | |
| 336 | 0.411/0.277 | 0.433/0.283 | 0.446/0.258 | 0.420/0.305 | |
| 720 | 0.445/0.294 | 0.467/0.302 | 0.476/0.265 | 0.459/0.323 | |
| Avg | 0.413/0.281 | 0.428/0.282 | 0.448/0.261 | 0.419/0.303 |
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.
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.
| Dataset | Seed 2024 | Seed 2025 | Seed 2026 | Mean ± Std |
|---|---|---|---|---|
| Metric | MSE/MAE | MSE/MAE | MSE/MAE | MSE/MAE |
| ETTh1 | 0.414/0.430 | 0.420/0.435 | 0.409/0.428 | 0.4143±0.0055 / 0.4310±0.0036 |
| ETTh2 | 0.344/0.392 | 0.350/0.397 | 0.341/0.390 | 0.3450±0.0046 / 0.3930±0.0036 |
| ETTm1 | 0.356/0.377 | 0.361/0.381 | 0.354/0.375 | 0.3570±0.0036 / 0.3777±0.0031 |
| ETTm2 | 0.254/0.313 | 0.258/0.316 | 0.252/0.312 | 0.2547±0.0031 / 0.3137±0.0021 |
| Weather | 0.225/0.256 | 0.230/0.260 | 0.223/0.254 | 0.2260±0.0036 / 0.2567±0.0031 |
| Electricity | 0.161/0.254 | 0.164/0.258 | 0.160/0.253 | 0.1617±0.0021 / 0.2550±0.0026 |
| Traffic | 0.413/0.281 | 0.421/0.286 | 0.409/0.279 | 0.4143±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.
| Model | Total Parameters | BLIP-2 Parameters | Inference Time (ms/batch) | Peak Memory (MiB) |
|---|---|---|---|---|
| EviTime | 3,791,113,705 | 3,744,761,856 | 613.316 | 19,243.118 |
| Time-VLM | 3,745,012,870 | 3,744,761,856 | 595.369 | 19,502.126 |
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).
| Dataset | EviTime (Pretrained) | EviTime (Random Init.) |
|---|---|---|
| Metric | MSE/MAE | MSE/MAE |
| ETTh1 | 0.365/0.391 | 0.402/0.419 |
| ETTh2 | 0.272/0.337 | 0.331/0.373 |
| ETTm1 | 0.298/0.343 | 0.347/0.381 |
| ETTm2 | 0.163/0.253 | 0.185/0.273 |
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.
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.