Organizations: College of Control Science and Engineering, Zhejiang University, Hangzhou, China · Zhejiang Provincial Security Science and Technology Research Institute, Hangzhou, China · Zhejiang Province Key Laboratory of “Public Security Brain” Big Data and Key Security Technologies Research, Hangzhou, China
Long-term time-series forecasters typically optimize a single numerical representation, even though the same history exposes complementary geometry, recurrence, and semantic structure. We propose AnchorMTS, which derives self-aligned visual and textual views from each observed window and uses them only to refine a complete numerical forecast. A trainable evidence path produces a candidate forecast, while a parameter-free output projection enforces an exact per-variable bound on its displacement from the numerical anchor. Across six benchmarks and four horizons, AnchorMTS achieves the best macro MSE/MAE (0.336/0.367) among seven supervised baselines and wins 18 of 24 MSE settings against the cellwise baseline envelope; removing the auxiliary path raises macro MSE to 0.340. The projection leaves 95.8% of candidate refinements unchanged while yielding zero bound violations, showing that complementary multi-view evidence can improve forecasting while its final influence remains explicitly controlled.
Figures & tables
Figure 1: Overview of AnchorMTS. The observed history is processed by a complete numerical forecaster to produce the anchor prediction, while self-derived angular, trajectory, recurrence, and textual views provide cached auxiliary evidence. Evidence fusion proposes a multimodal candidate, and an exact forecast-space projection converts its residual into a bounded refinement of the numerical anchor.
Figure 2: Forecasting accuracy. (a) Macro MSE over the 24 aligned dataset-horizon cells. (b) Dataset-average MSE difference from TimePro, the strongest individual baseline by macro error; negative values favor AnchorMTS.
Method
MSE ↓
MAE ↓
MSE W/T/L
MAE W/T/L
AnchorMTS
0.336
0.367
–
–
TimePro
0.348
0.374
22/0/2
20/1/3
TimeXer
0.351
0.374
20/1/3
18/1/5
TimeMixer
0.358
0.384
23/0/1
21/1/2
iTransformer
0.358
0.379
24/0/0
23/1/0
PatchTST
0.372
0.388
23/0/1
23/0/1
Table 1: Macro comparison over 24 aligned dataset–horizon settings. W/T/L compares AnchorMTS with each baseline at the reported three-decimal precision. *Cellwise best denotes the metric-wise envelope over the seven baselines rather than an additional forecasting model.
Figure 3: Exact forecast-space projection. (a) Correction-ratio distributions before and after the final projection over all 24 settings; candidate refinements inside ρf=0.25 pass unchanged, while the extreme tail is truncated at the prescribed radius. (b) Fraction of sample–variable forecasts projected in each dataset. Overall, 95.8% of refinements pass unchanged and no reported forecast violates the bound.
Frozen view
Trend
Period.
Recur.
Chan. div.
Var.-wise GAF
0.107
0.109
0.109
0.161
Joint line
0.279
0.324
0.329
0.039
Recurrence plot
0.259
0.248
0.287
0.042
Table 2: Retrieval lift of frozen visual representations, macro-averaged across six training sets. Higher values indicate that representation-space neighbors are more similar in the corresponding numerical signal descriptor than arbitrary window pairs.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
H
AnchorMTS
TimePro
TimeXer
TimeMixer
iTransformer
PatchTST
TimesNet
DLinear
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
ETTh1
96
0.382
0.401
0.375
0.398
0.382
0.403
0.385
0.402
0.386
0.405
0.460
0.447
0.384
0.402
0.395
0.407
192
0.430
0.430
0.427
0.429
0.429
0.435
0.443
0.430
0.441
0.436
0.512
0.477
0.436
0.429
0.446
0.441
336
0.463
0.445
0.472
0.450
0.468
0.448
0.512
0.470
0.487
0.458
0.546
0.496
0.491
0.469
0.489
0.467
720
0.471
0.470
0.476
0.474
0.469
0.461
0.497
0.476
0.503
0.491
0.544
0.517
0.521
0.500
0.513
0.510
ETTh2
96
0.280
0.331
0.293
0.345
0.286
0.338
0.289
0.342
0.297
0.349
0.308
0.355
0.340
0.374
0.340
0.394
Appendix
Table 3: Detailed dataset–horizon comparison against all seven Table 1 baselines. Bold values mark the lowest reported three-decimal error in each row and metric, so MSE and MAE winners may come from different methods. TimeXer ETT/Weather values follow its original report, while its Exchange row follows the aligned input-length-96 reproduction of [ 23 ] ; other external values follow their recorded aligned sources.
Dataset
Horizon
AnchorMTS
TimePro
Best MSE
Best MAE
ETTh1
96
0.382/0.401
0.375/0.398
TimePro 0.375
TimePro 0.398
ETTh1
192
0.430/0.430
0.427/0.429
TimePro 0.427
TimePro 0.429
ETTh1
336
0.463/0.445
0.472/0.450
TimeXer 0.468
TimeXer 0.448
ETTh1
720
0.471/0.470
0.476/0.474
TimeXer 0.469
TimeXer 0.461
ETTh2
96
0.280/0.331
0.293/0.345
TimeXer 0.286
TimeXer 0.338
ETTh2
192
0.357/0.382
0.367/0.394
TimeXer 0.363
TimeXer 0.389
Appendix
Table 4: Complete fixed-identity AnchorMTS grid and the cellwise best Table 1 references. Best MSE and Best MAE are selected independently from all seven displayed baselines in each cell.
Dataset
Horizon
d
dff
Layers
Batch
LR
Loss
ETTh1
96
512
2048
2
32
1.0e-04
Huber
ETTh1
192
512
2048
2
32
1.0e-04
Huber
ETTh1
336
512
2048
2
32
1.0e-04
Huber
ETTh1
720
128
512
1
16
1.0e-04
MSE
ETTh2
96
512
2048
2
32
1.0e-04
Huber
ETTh2
192
512
2048
2
32
1.0e-04
Huber
Appendix
Table 5: Selected ordinary training and temporal-capacity hyperparameters. Every row retains the identical multimodal protocol.
Default look-back 336; Exchange absent; no complete aligned 6×4 grid.
Alternative-view fusion context
TimesCLIP [ 13 ]
Supervised aligned views
Input length 96, but ETTh1/ETTh2 and horizons 336/720 are not reported.
Cross-view alignment context
Appendix
Table 6: Published evaluation protocols of the closest multimodal forecasters. Their reported grids differ from Table 1 in context length, dataset coverage, and/or horizon coverage; partial results are therefore not mixed into the 24-setting macro comparison.
Dataset
Variables
Frequency
Split
ETTh1/ETTh2
7
hourly
12/4/4 months
ETTm1/ETTm2
7
15-minute
12/4/4 months
Exchange
8
daily
70/10/20%
Weather
21
10-minute
70/10/20%
Appendix
Table 7: Dataset metadata and chronological train/validation/test partitions. Input length is 96 and horizons are 96, 192, 336, and 720.
Figure 4: Representations generated from one Exchange test history. The variable-wise GAF preserves channel-specific pairwise geometry; line and recurrence views expose trajectory and return structure; the caption summarizes the same observed window. No target-window values are used. The GAF is shown in grayscale for display; caption generation uses the corresponding color rendering.
Web platforms run on multivariate time series, from service telemetry and Web traffic logs to sensor streams served through Web APIs. Yet, retention policies often archive them at coarse granularity to save storage and bandwidth. Decisions such as autoscaling, alerting, and billing operate at finer granularity than existing forecasters reach, since these models predict at the resolution of their input. We therefore introduce Multivariate Time Series Super-Resolution Forecasting (MTS-SRF), which forecasts fine-grained futures from coarse histories, and propose TimeMagnifier to address its three challenges. Fourier Anchor counters data sparsity by fitting each input window to a trend-Fourier structure with sparse cross-variable gating; De-noise Only constrains fine-grained ambiguity by refining along the resolution axis rather than the noise axis, lengthening the grid level by level; and banded DTW Alignment corrects cross-resolution misalignment against the coarse forecast. On the first MTS-SRF benchmark, with ten datasets from five domains, real archived cross-resolution pairs, strengthened baseline adaptations, and five-seed significance tests, TimeMagnifier is best on every dataset, resolution coefficient, and metric against 15 baselines and cuts MSE by 12.3% on average. Its De-noise Only and DTW stages also transfer as a plug-in that lowers the MSE of existing forecasters by 5.1% to 9.0%. Code will be made publicly available.
Shuhao Li, Weidong Yang, Yue Cui +2
Fudan University, Shanghai, China · Nanyang Technological University, Singapore · Zhuhai Fudan Innovation Research Institute, Zhuhai, China +3
Long-term forecasting models commonly process all patches in a look-back window using the same fixed stack. Older contextual patches and recent evidence therefore receive the same computational depth. Yet the information closest to the forecast and the more distant context do not contribute equally. Uniform processing leaves this distinction unexpressed in the architecture. We propose MARO, a Most-Recent Anchoring with Recurrent Ordering model that processes the look-back window from the most recent patch to the oldest. The most recent patch serves as the anchor. It initializes the latent state and conditions each subsequent step, so older patches are folded into a representation that remains centered on recent evidence. A single shared module is reused at every step, so extending the scan further into the past introduces no additional parameters. Intermediate states retained during the scan allow the forecast head to weigh short and long portions of the history separately. This expresses recency through the order of recurrent refinement. Extensive experiments across multiple real-world time series datasets show that MARO achieves state-of-the-art performance on both long-term and short-term forecasting tasks.Ablation studies examine the contribution of the main architectural components.
Jung Min Choi, Ngoc Son Le, Ibram Abdelmalak +2
VWFS Data Analytics Research Center (VWFS-DARC), Hildesheim, Germany
Multivariate time-series forecasting is essential to many real-world applications. Recent large vision models (LVMs) offer a promising paradigm by transferring cross-domain visual priors to time-series forecasting. However, existing LVM-based methods face two key challenges: balancing independent visual representation spaces with cross-variable dependency modeling, and adapting vision backbones pretrained on natural images to the distinct temporal semantics of time-series images. To address these challenges, we propose MUSE, a dependency-aware adaptation framework built on a fully frozen pretrained MAE. First, the Variable Context Refinement Module (VCR) aggregates shared temporal information within each variable and models cross-variable contextual dependencies while preserving independent visual spaces. Second, the Temporal-Periodic Refinement Module (TPR) performs lightweight refinement at different encoder depths and explicitly models across-period temporal dependencies and within-period periodic dependencies. The two modules independently produce forecasts, which are fused through a learnable prediction-level gate. Experiments on 10 real-world datasets demonstrate that MUSE achieves state-of-the-art performance.