Organizations: Institute of Computing Technology, Chinese Academy of Sciences · Beijing University of Posts and Telecommunications · University of Chinese Academy of Sciences · Northeastern University · Zhejiang Lab
Time series forecasting underpins critical applications in finance, energy, healthcare, and transportation. Although deep models have achieved strong results, most adopt single-scale modeling or restrict multiscale processing to the input side, causing a misalignment between multiscale inputs and single-scale outputs and limiting predictive power. We introduce the Modular Scale-wise Autoregressive Framework (MSAR), a model-agnostic design that forecasts progressively across multiple temporal resolutions. MSAR offers three advantages: (1) scale-wise aligned modeling, which disentangles heterogeneous temporal patterns by aligning inputs and outputs at each scale; (2) scale-wise autoregression, where coarse-scale predictions guide finer-scale forecasting through hierarchical information flow; and (3) a modular architecture, enabling seamless integration with diverse backbones such as CNNs, MLPs, and Transformers. Extensive experiments across a broad set of datasets and forecasting models demonstrate that MSAR achieves consistent improvements in both accuracy and inference efficiency, validating the effectiveness of scale-aligned autoregression for multiscale time series forecasting. The code is available at: https://github.com/Oryxo/MSAR.
Figures & tables
Figure 1: Comparison of forecasting paradigms. MSAR introduces scale-wise autoregression with aligned multi-scale inputs and outputs. Coarse-scale predictions serve as contextual guidance for finer-scale forecasting through hierarchical information flow.
Figure 2: Overview of the MSAR framework. MSAR is a model-agnostic , scale-wise autoregressive pipeline over multi-resolution inputs obtained by stride down-sampling. The pipeline starts from the coarsest scale and progressively proceeds to finer scales, where each module concatenates current-scale inputs with aligned coarse predictions and performs imputation-based refinement . Through this progressive imputation process, MSAR constructs scale-aligned inputs and produces coarse-to-fine multi-scale forecasts .
SimpleTM
DUET
iTransformer
PatchTST
DLinear
Dataset
Metric
Base
+MSAR
Base
+MSAR
Base
+MSAR
Base
+MSAR
Base
+MSAR
ECL
MSE
0.164
0.168
0.178
0.177
0.167
0.160
0.168
0.166
0.170
0.168
MAE
0.254
0.259
0.277
0.276
0.257
0.252
0.258
0.254
0.269
0.259
ETTh1
MSE
0.424
0.410
0.403
0.404
0.475
0.449
0.428
0.413
0.420
0.419
MAE
0.437
0.425
0.415
0.416
0.459
0.455
0.439
0.428
0.430
0.426
ETTh2
MSE
0.377
0.351
0.355
0.345
0.405
0.395
0.404
0.348
0.397
0.391
Table 2: Long-term forecasting with input length 336, averaged over horizons {96,192,336,720} . Lower is better; bold marks pairwise minima at the reported precision, including ties. Full results: Appendix B .
Figure 4
Figure 4: Impact of look-back window length on forecasting performance (measured by MSE). We report results on three representative datasets (ETTm1, Weather, and Electricity) under two horizons ( H=96 and H=720 ). Each curve shows the error as the input sequence length increases from 96 to 1152 .
Figure 5: MSAR sensitivity on ETTh1; lower is better. (a) Interval lists (s,1) , averaged over five backbones. (b) Layer counts for PatchTST and SimpleTM; one layer is the baseline. Axis limits are shared within each metric.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Dim
Dataset Size (Train/Val/Test)
Frequency
ETTh1, ETTh2
7
(8545, 2881, 2881)
Hourly
ETTm1, ETTm2
7
(34465, 11521, 11521)
15 min
Weather
21
(36792, 5271, 10540)
10 min
Traffic
862
(12185, 1757, 3509)
Hourly
Electricity
321
(18317, 2633, 5261)
Hourly
Exchange-rate
8
(5120, 665, 1422)
Daily
Appendix
Table 4: Dataset statistics. “Dim” denotes the number of variates; “Dataset Size” reports benchmark (Train/Validation/Test) split sizes; “Frequency” is the sampling interval. PEMS split sizes follow iTransformer [ 12 ] .
Dataset / Configurations
Model Hyper-parameter
Training Process
dmodel
dff
Layers
Heads
Batch Size
Epochs
LR ∗
Patience
ECL
128
256
1
2
16
10
10−3
3
Traffic
128
256
1
2
16
10
10−3
3
Solar
128
256
1
2
16
10
10−3
3
ETTh1
512
2048
2
8
32
10
10−3
3
ETTh2
512
2048
2
8
32
10
10−3
3
Appendix
Table 5: Experiment configuration of MSAR in Table 6 . All experiments use the Adam optimizer with a learning rate of 10−3 and early stopping (patience = 3).
Models
Metric
SimpleTM
DUET
iTransformer
PatchTST
DLinear
Base
+Ours
Base
+Ours
Base
+Ours
Base
+Ours
Base
+Ours
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
Electricity
96
0.134
0.226
0.137
0.231
0.149
0.251
0.147
0.250
0.135
0.229
0.132
0.226
0.138
0.231
0.136
0.227
0.143
0.242
0.141
0.234
192
0.154
0.245
0.152
0.244
0.163
0.264
0.162
0.263
0.153
0.245
0.149
0.241
0.154
0.246
0.152
0.241
0.157
0.255
0.155
0.246
336
0.175
0.268
0.167
0.258
0.181
0.281
0.179
0.279
0.169
0.261
0.164
0.257
0.171
0.261
0.168
0.257
0.172
0.272
0.170
0.262
720
0.191
0.278
0.214
0.303
0.220
0.311
0.219
0.310
0.209
0.294
0.196
0.285
0.210
0.294
0.207
0.290
0.207
0.305
0.204
0.293
Appendix
Table 6: Full long-term forecasting results on 10 datasets. Avg rows average the four displayed horizons, rounded to three decimals; ties at this precision are both bold.
Models
MSAR
SimpleTM
DUET
iTransformer
PatchTST
DLinear
Ours
[ 2025 ]
[ 2025 ]
[ 2024a ]
[ 2023 ]
[ 2023 ]
Metric
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
ETTm1
96
0.289
0.335
0.280
0.337
0.300
0.345
0.299
0.348
0.289
0.343
0.299
0.343
192
0.325
0.355
0.326
0.365
0.335
0.366
0.334
0.373
0.332
0.367
0.336
0.367
336
0.355
0.376
0.362
0.387
0.364
0.383
0.370
0.394
0.362
0.389
0.367
0.385
720
0.410
0.406
0.417
0.422
0.417
0.413
0.422
0.426
0.408
0.420
0.420
0.418
Appendix
Table 7: Full hyperparameter search results on five datasets with prediction lengths {96,192,336,720} . Each model is evaluated with its best configuration within the shared search space; the look-back window is selected from {96,192,320,512} . Avg. denotes the average across prediction lengths. Search details are provided in Appendix A.3 .
Dataset
Metric
SimpleTM
DUET
iTransformer
PatchTST
DLinear
Base
+Ours
Base
+Ours
Base
+Ours
Base
+Ours
Base
+Ours
PEMS03
MAE
19.147
15.627
18.853
18.402
15.870
15.308
16.974
15.209
17.411
16.743
MAPE
19.527
15.796
22.031
20.698
15.945
15.459
19.376
15.905
15.876
15.509
RMSE
30.307
24.933
29.607
29.240
25.563
24.648
26.517
24.310
29.619
27.909
PEMS04
MAE
27.192
21.651
24.581
24.693
22.033
20.617
23.005
21.022
24.359
22.546
MAPE
18.064
13.462
18.659
18.202
14.037
13.019
15.779
13.846
14.616
13.538
Appendix
Table 8: Short-term forecasting results on PEMS datasets (prediction length = 12). Lower MAE, MAPE, and RMSE indicate better performance.
Figure 12
Figure 7: Visualization of model predictions and ground truth on the ETTh1 dataset. Results are produced by the DUET+MSAR model.
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
Time series forecasting is vital across many domains, yet existing models struggle with fixed-length inputs and inadequate multi-scale modeling. We propose MR-CDM, a framework combining hierarchical multi-resolution trend decomposition, an adaptive embedding mechanism for variable-length inputs, and a multi-scale conditional diffusion process. Evaluations on four real-world datasets demonstrate that MR-CDM significantly outperforms state-of-the-art baselines (e.g., CSDI, Informer), reducing MAE and RMSE by approximately 6-10 to a certain degree.
Xianyong Xu, Yuanjun Zuo, Zhihong Huang +4
Department, State Grid Hunan Electric Power Company Limited Research Institute & Hunan Province Engineering Technology Research Center of Electric Power Multimodal Perception and Edge Intelligence, Changsha, Hunan, China. · College of Computer Science, Hunan University, Changsha, Hunan, China.
Multi-scale modeling has emerged as an effective design principle for time-series forecasting by capturing temporal dynamics at multiple resolutions. As no principled foundation has been established in the literature, we unify existing scaling methods into a scaling operator family, revealing a fundamental limitation of existing approaches: reliance on fixed and discrete scaling. To address this limitation, we propose SiGMA (Single Generalized Multi-scale Architecture), which enables distance-aware scaling via the learnable discrete Gaussian (LDG) kernel grounded in scale-space theory. We evaluate SiGMA comprehensively on long- and short-term forecasting benchmarks against state-of-the-art multi-scale baselines. SiGMA outperforms all competitors on both tasks, especially achieving the best performance in 13 out of 16 long-term evaluation settings. Beyond accuracy, SiGMA significantly improves training speed by up to 5.3 times and reduces memory consumption by up to 3.8 times over the strongest competitors. Code is available at https://github.com/cheonwoolee/SiGMA.
Cheonwoo Lee, Dooho Lee, Doyun Choi +1
School of Electrical Engineering, KAIST, Daejeon, Republic of Korea · Department of Computer Science and Engineering, Seoul National University, Seoul, Republic of Korea
Time Series Forecasting (TSF) faces persistent challenges in modeling intricate temporal dependencies across different scales. Despite recent advances leveraging different decomposition operations and novel architectures based on CNN, MLP or Transformer, existing methods still struggle with static decomposition strategies, fragmented dependency modeling, and inflexible fusion mechanisms, limiting their ability to model intricate temporal dependencies. To explicitly solve the mentioned three problems respectively, we propose a novel Dynamic Multi-Scale Coordination Framework (DMSC) with Multi-Scale Patch Decomposition block (EMPD), Triad Interaction Block (TIB) and Adaptive Scale Routing MoE block (ASR-MoE). Specifically, EMPD is designed as a built-in component to dynamically segment sequences into hierarchical patches with exponentially scaled granularities, eliminating predefined scale constraints through input-adaptive patch adjustment. TIB then jointly models intra-patch, inter-patch, and cross-variable dependencies within each layer's decomposed representations. EMPD and TIB are jointly integrated into layers forming a multi-layer progressive cascade architecture, where coarse-grained representations from earlier layers adaptively guide fine-grained feature extraction in subsequent layers via gated pathways. And ASR-MoE dynamically fuses multi-scale predictions by leveraging specialized global and local experts with temporal-aware weighting. Comprehensive experiments on thirteen real-world benchmarks demonstrate that DMSC consistently maintains state-of-the-art (SOTA) performance and superior computational efficiency for TSF tasks. Code is available at https://github.com/1327679995/DMSC.
Haonan Yang, Jianchao Tang, Zhuo Li +1
National University of Defense Technology, Changsha, Hunan, China