betaFeedback

WorksComputer Science

MSAR: Next-Scale Autoregressive Forecasting for Time Series via Modular Multi-Scale Decoupling

Abstract

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.

Introduction

Time series forecasting is critical in domains such as finance, energy, healthcare, and transportation. Recent deep models—including RNNs [1, 2, 3, 4], CNNs [5, 6], MLPs [7, 8, 9, 10], and Transformers [11, 12, 13]—have shown strong performance. However, they typically rely on single-scale modeling, processing historical data at a fixed temporal scale. This overlooks a key property of real-world time series: temporal patterns vary across scales [14, 15]. Forcing heterogeneous patterns into a unified representation often causes scale interference, leading to poor generalization and degraded forecasting accuracy.

To address the limitations of single-scale modeling, prior works [16, 17, 18, 14, 19, 20, 21, 22, 23] introduced multiscale-mixing in input modeling to capture patterns at different scales. However, they still predict future values at a single (fine) scale, leading to a mismatch between multiscale inputs and single-scale outputs. Similarly, multi-resolution decoders [24, 25] generate coarse-to-fine intermediate components, but these are ultimately interpolated into a single-resolution prediction with no per-scale supervision. This misalignment complicates the learning process and weakens the model’s ability to leverage multiscale information effectively, often resulting in suboptimal forecasts.

Building upon the aforementioned limitations, we propose the Modular Scale-wise Autoregressive Framework (MSAR). As shown in Figure 1, MSAR performs scale-aligned and decoupled modeling, where each scale is modeled independently. This explicit alignment helps isolate heterogeneous temporal patterns and simplifies the learning process. Motivated by Tian et al. [ [26], we introduce a scale-wise autoregressive forecasting strategy. This structured design clearly distinguishes MSAR from prior approaches that rely on single-scale decoding [27, 28, 29, 12, 11, 8], from cross-scale mixing without resolution consistency [23, 22, 14, 20], and from multiresolution decoders [24, 25]. In contrast, MSAR enforces explicit per-scale supervision and strict within-scale alignment, enabling principled hierarchical autoregressive refinement across scales.

Comparison of forecasting paradigms

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.

Specifically, MSAR adopts a scale-wise modular framework that enables flexible implementation across different forecasting models. This design makes the MSAR model-agnostic and easy to integrate into a variety of backbone architectures. To coordinate the prediction process across scales, MSAR introduces a simple yet effective information flow mechanism, where predictions from coarser scales serve as contextual signals to guide finer-scale forecasting. Furthermore, to mitigate potential errors propagated from earlier predictions, we incorporate a lightweight refinement module that selectively fuses coarse predictions with current-scale inputs. This refinement process improves prediction accuracy and enhances consistency across scales.

MSAR offers three key advantages over prior approaches:

  • Scale-wise aligned modeling: Unlike previous methods that apply multiscale-mixing [23, 22, 14, 20] only to input features, MSAR performs fully separated modeling at each scale, with inputs and outputs aligned within the same scale. This design reduces learning complexity by isolating heterogeneous temporal patterns and avoiding interference across scales.

  • Scale-wise autoregression: MSAR adopts a scale-wise autoregressive strategy [26], where predictions from coarser scales serve as contextual signals for finer-scale forecasting. This hierarchical structure improves the model’s ability to capture long-term dependencies while enabling progressive refinement of short-term patterns, thereby enhancing both accuracy and interpretability.

  • Modular and model-agnostic design: Each scale in MSAR is processed by an independent forecasting module without architectural constraints. This modular design enables seamless integration with diverse backbone models—such as CNNs [5], MLPs [8], and Transformers [11, 12], and allows flexible adaptation to varying input lengths.

Related Work

Time Series Forecasting

Single-Scale Modeling. Modern time series forecasting is predominantly driven by deep learning models, including RNN-based [4], CNN-based [5], Transformer-based [11, 12], and MLP-based [8], 29] architectures. However, these methods typically operate at a single temporal scale and fail to capture distinct patterns that emerge across different resolutions. As a result, long-context inputs often lead to overfitting, as the model attempts to fit both global trends and local fluctuations within a unified representation, without scale-aware separation [28].

Multiscale Modeling. To capture temporal dynamics across different frequencies, several recent works have introduced multiscale modeling into time series forecasting. SimpleTM [23] employs wavelet-based tokenization to generate scale-specific representations for Transformers. RAFT [22] leverages retrieval across different periodicities to inject external multi-scale patterns into forecasting. TimeMixer [14] employs grouped mixing layers to model dependencies across temporal resolutions in a unified encoder. Pathformer [20] constructs adaptive pathways to route information across temporal resolutions and distances. These methods effectively encode multiscale patterns in the input sequence but still predict at a single temporal scale, creating a mismatch between input and output resolutions. In contrast, our MSAR performs scale-aligned, coarse-to-fine autoregressive decoding: each stage only consumes inputs at its assigned resolution and is causally conditioned on the previous (coarser) stage, enabling consistent multiscale modeling across both input and output without fusion-based mismatch.

Multi-Resolution Decoders. Many recent architectures employ multi-resolution decoders, where intermediate temporal scales are generated or refined before producing the final forecast. Crossformer [18] and MR-Transformer [30] construct hierarchical or segment-level multi-resolution pathways, yet only the final fine-scale output is supervised. Scaleformer [19] performs iterative multi-scale decoding, but all intermediate resolutions are fused into a single prediction through pooling or learned interpolation. xLSTM-Mixer [25] mixes multi-span temporal memories inside a multi-resolution decoding stack but still emits a single-resolution forecast. N-HiTS [24] explicitly reconstructs intermediate blocks at multiple resolutions, but these are subsequently upsampled and aggregated rather than decoded as separate supervised outputs. MuSiCNet [31] introduces a gradual coarse-to-fine framework for irregularly sampled multivariate time series, progressively refining latent representations to handle temporal sparsity. Thus, although these methods use multi-resolution decoding mechanisms, none of them performs explicit per-scale decoding with per-scale targets.

Autoregressive Approaches for Sequence Modeling

Autoregressive (AR) modeling has been a widely adopted paradigm in time series forecasting. DeepAR [1] and SutraNets [3] generate the future sequence one time step at a time, where each prediction depends on historical inputs and previously generated values. SegRNN [4] extends conventional RNNs with segment-wise recurrence and parallel multi-step decoding. C2FAR [2] adopts a hierarchical binning strategy to generate each value autoregressively from coarse to fine levels of numerical precision. Despite their differences, all these methods operate at a single temporal scale, which limits their ability to capture complex temporal dependencies that span across multiple resolutions. Recent time-series foundation models such as Time-MoE [27] and Timer-XL [28] also adopt autoregression over time steps, leveraging large model capacity and extensive pretraining to achieve strong performance. However, their focus is on scaling parameters and training data, whereas our work centers on paradigm improvements.

Motivated by VAR [26], we propose the MSAR, which formulates forecasting as a next-scale prediction task across scale-aligned temporal resolutions. MSAR takes as input a historical time series and decomposes both the inputs and prediction targets into multiple aligned temporal scales. Each scale is modeled separately using a plug-in forecasting module, and predictions from coarser scales are progressively used to condition finer-scale forecasts, enabling a hierarchical flow of information.

Methodology

Problem Formulation Let the time series data be represented as: X1:T={x1,x2,…,xT}∈RT×C\mathbf{X}_{1:T} = \{\mathbf{x}_1, \mathbf{x}_2, \ldots, \mathbf{x}_T\} \in\mathbb{R}^{T \times C}, where TT denotes the length of the look-back window and CC represents the number of input channels. The goal is to predict the future sequence: Y1:H∈RH×C\mathbf{Y}_{1:H} \in\mathbb{R}^{H \times C}, where HH is the forecast horizon. For simplicity, we directly use X\mathbf{X} and Y\mathbf{Y} to represent the input and output data, respectively.

Overview of the MSAR framework

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.

Notations We define X1:TΔ∈R⌊T/Δ⌋×CX^{\Delta}_{1:T} \in\mathbb{R}^{\lfloor T/\Delta\rfloor\times C} as the downsampled sequence over time steps, where ⌊T/Δ⌋\lfloor T/\Delta\rfloor represents the dimension after downsampling. Similarly, we define Y1:HΔ∈R⌊H/Δ⌋×CY^{\Delta}_{1:H} \in\mathbb{R}^{\lfloor H/\Delta\rfloor\times C} as the prediction results with a time interval of Δ\Delta.

Technical Motivation

To understand how scale alignment and temporal granularity affect forecasting performance, we conduct experiments on the ETTh1 dataset with a fixed prediction horizon of H=336H = 336. As shown in Table 1, we evaluate models under three input settings: (1) full-resolution input (T=576T = 576); (2) downsampled3 input with full-resolution output (T=576Δ=6,H=336T = 576^{\Delta=6}, H = 336); and (3) fully scale-aligned input-output (T=576Δ=6,H=336Δ=6T = 576^{\Delta=6}, H = 336^{\Delta=6}). In addition to standard MSE, we report MSEΔ=6\mathrm{MSE}^{\Delta=6}, which measures prediction quality under 6× temporal downsampling.

ModelTHMSEMSEΔ=6\mathrm{MSE}^{\Delta=6}
PatchTST5763360.4800.477
PatchTST576Δ=6576^{\Delta=6}3360.5040.445
PatchTST576336Δ=6336^{\Delta=6}-0.564
PatchTST576Δ=6576^{\Delta=6}336Δ=6336^{\Delta=6}-0.433
DLinear5763360.4450.449
DLinear576Δ=6576^{\Delta=6}3360.4650.438
DLinear576336Δ=6336^{\Delta=6}-0.457
DLinear576Δ=6576^{\Delta=6}336Δ=6336^{\Delta=6}-0.425
TimeMixer5763360.4320.435
TimeMixer576336Δ=6336^{\Delta=6}-0.461

Table 1. Forecasting results on the ETTh1 dataset. MSEΔ=6\mathrm{MSE}^{\Delta=6} reflects the MSE under 6× downsampling. 576Δ=6576^{\Delta=6} and 336Δ=6336^{\Delta=6} correspond to context and prediction sequences with 6× downsampling.

We highlight three key observations:

❶ Scale misalignment degrades performance. When input and output scales are not aligned (e.g., T=576Δ=6,H=336T = 576^{\Delta=6}, H = 336), performance consistently drops compared to the fully aligned setting. This degradation occurs despite the downsampled input offering more context in real time steps, underscoring the importance of maintaining resolution consistency between input and output.

❷ Sparse, scale-aligned forecasting improves coarse-scale accuracy. Models trained and evaluated under the same coarse resolution (T=576Δ=6,H=336Δ=6T = 576^{\Delta=6}, H = 336^{\Delta=6}) achieve the lowest MSEΔ=6\mathrm{MSE}^{\Delta=6}, demonstrating that explicitly modeling forecasting at sparse temporal scales yields more accurate coarse-level predictions—even with significantly fewer input tokens.

③ Multiscale-mixing improves overall accuracy but lacks scale-specific precision. TimeMixer [14], which adopts a multiscale-mixing strategy through grouped temporal operations, achieves the lowest overall MSE. However, its coarse-scale performance (as indicated by MSEΔ=6\mathrm{MSE}^{\Delta=6}) remains inferior to that of models trained explicitly at coarse resolutions. This suggests that while multiscale-mixing enhances global modeling capacity, it may obscure scale-specific patterns and reduce fidelity at individual temporal granularities.

These findings motivate our proposed framework, which models forecasting at multiple aligned temporal scales, aiming to preserve both temporal coherence and resolution-specific expressiveness.

Modular Scale-wise Autoregressive Framework

Real-world time series often exhibit temporal patterns at multiple resolutions—long-term trends unfold gradually, while short-term fluctuations vary rapidly. Modeling such heterogeneous dynamics using a fixed-resolution representation often leads to underfitting or overfitting. To address this, we propose the Modular Scale-wise Autoregressive Framework (MSAR), a model-agnostic pipeline that decomposes forecasting into multiple scale-aligned subproblems. Each scale is handled independently by a dedicated forecasting module, while a lightweight information flow mechanism propagates relevant context from coarse to fine resolutions.

Scale-wise Decoupling and Progressive Forecasting. We define a sequence of temporal scales with downsampling factors {d1,d2,…,dN}\{d_1,d_2,\ldots,d_N\}, where d1>d2>⋯>dN=1d_1>d_2>\cdots>d_N=1. Given an input sequence XT−W:T∈RW×CX_{T-W:T}\in\mathbb{R}^{W\times C} and a forecasting horizon HH, the scale-ii input and target are

Xi=XT−Wi:TΔ=di∈RWidi×C,Yi=Y1:HΔ=di∈RHdi×C,(1)X^i=X_{T-W_i:T}^{\Delta=d_i}\in\mathbb{R}^{\frac{W_i}{d_i}\times C},\quad Y^i=Y_{1:H}^{\Delta=d_i}\in\mathbb{R}^{\frac{H}{d_i}\times C}, \tag*{(1)}

where XΔ=dX^{\Delta=d} denotes a deterministic stride-based sampling operator that retains every dd-th timestamp without interpolation or filtering, thereby preserving the original temporal structure at a coarser resolution. The WiW_i is the look-back window at scale ii.

Forecasting Objective. MSAR formulates multiscale forecasting as a structured autoregressive process:

p(Y1,…,YN∣X1,…,XN)=∏i=1Np(Yi∣Xi,Y^align<i),(2)p(Y^1,\ldots,Y^N\mid X^1,\ldots,X^N)=\prod_{i=1}^{N}p(Y^i\mid X^i,\widehat{Y}^{<i}_{\mathrm{align}}), \tag*{(2)}

where Y^align<i\widehat{Y}^{<i}_{\mathrm{align}} aggregates coarse-scale predictions whose time indices coincide with those of scale ii:

Y^align<i=⋃j<i{Y^tj∣t∈τj∩τi}∈RHdi×C,(3)\widehat{Y}^{<i}_{\mathrm{align}}=\bigcup_{j<i}\left\{\widehat{Y}^{j}_{t}\mid t\in\tau^j\cap\tau^i\right\}\in\mathbb{R}^{\frac{H}{d_i}\times C}, \tag*{(3)}

with τj\tau^j the index set at scale jj. For time points in τi\tau^i that receive no coarse-scale coverage, we insert zero vectors to match the target size of YiY^i:

t∉⋃j<iτj⇒Y^align<i[t]=0.(4)t\notin\bigcup_{j<i}\tau^j\quad\Rightarrow\quad\widehat{Y}^{<i}_{\mathrm{align}}[t]=0. \tag*{(4)}

Information Flow Mechanism. The role of each prediction module Fθi\mathcal{F}^i_{\theta} differs by scale. At the coarsest level (i=0i=0), Fθ0\mathcal{F}^0_{\theta} acts as a forecasting model, generating an initial coarse-resolution prediction solely from the input X0X^0. For higher scales (i>0i>0), the modules operate as imputation models, where each Fθi\mathcal{F}^i_{\theta} refines missing fine-scale values by conditioning on both the current-scale input XiX^i and the aligned coarse predictions passed from previous stages. Formally,

Y^i=Fθi(concat⁡(Xi,Y^align<i)),Y^i∈RHdi×C.(5)\widehat{Y}^i=\mathcal{F}^i_{\theta}\left(\operatorname{concat}\left(X^i,\widehat{Y}^{<i}_{\mathrm{align}}\right)\right),\quad\widehat{Y}^i\in\mathbb{R}^{\frac{H}{d_i}\times C}. \tag*{(5)}

This hierarchical design enforces a scale-aligned information flow: coarse predictions provide structural guidance, while finer-scale modules progressively enhance temporal resolution through imputation-based refinement.

Refinement in the Imputation Phase. To further enhance cross-scale consistency, MSAR performs complete prediction over the target window YiY^{i} during the imputation phase. Instead of restricting supervision to only the masked positions (mask = 0), we compute the loss across all target positions in the prediction window, including those that were previously filled (mask = 1). This approach enables iterative refinement, ensuring that coarse-scale predictions are smoothly integrated and corrected at finer resolutions, while maintaining temporal consistency across scales. This simple yet effective method adjusts previous predictions without introducing any additional parameters, helping to mitigate error accumulation across scales.

Training Strategy

We train MSAR using the scale-wise MAE loss: L=∑i=1NMAE⁡(Y^i,Yi)\mathcal{L} = \sum_{i=1}^{N} \operatorname{MAE}(\hat{Y}^{i}, Y^{i}). Directly optimizing this objective with fully autoregressive inputs can lead to slow convergence. During training, finer-scale modules receive ground-truth values from coarser scales, while at inference they receive model predictions. This discrepancy, commonly known as exposure bias, motivates our two-stage training strategy, which gradually transitions from ground-truth conditioning to inference-aligned prediction.

Teacher Forcing Training. In the first stage, each scale-specific module FθiF_{\theta}^{i} is trained independently using clean contextual signals. The coarsest forecaster (i=0i = 0) is optimized directly on ground truth. For finer scales (i>0i > 0), the model concatenates the ground-truth values of Yalign<iY_{\mathrm{align}}^{<i} from coarser levels with the current input. This teacher-forced approach stabilizes early learning and accelerates convergence by preventing error propagation across scales. Importantly, during this phase, the imputation models at finer scales use the ground-truth YY values (rather than previous predictions), ensuring the model learns directly from the real target values without introducing refinement.

Joint Training. In the second stage, we switch to inference-aligned inputs: each imputation module receives predictions from coarser levels rather than ground truth. Crucially, during this phase we compute the loss across the forecast horizon at each scale (i.e., all Hdi\frac{H}{d_i} steps), rather than restricting supervision to unobserved positions. This full-horizon supervision enables progressive refinement: predictions inherited from coarse scales can be corrected at finer resolutions, ensuring temporal coherence across scales. Thanks to its modular and scale-independent design, this curriculum allows MSAR to transition smoothly from robust pretraining under teacher forcing to deployment-consistent autoregressive inference.

Experiments

We conduct extensive experiments to evaluate the effectiveness of the proposed Modular Scale-wise Autoregressive Framework (MSAR). Our empirical study is designed to address the following key questions: (1). Model-agnostic gains: As a plugin method, can MSAR consistently improve mainstream state-of-the-art (SOTA) forecasting models, and how does it compare against strong baselines? (cf. § 4.1) (2). Efficiency study: What is the training and inference overhead introduced by MSAR compared to vanilla models? (cf. § 4.1) (3). Mechanism analysis: Why does MSAR work? We perform ablation studies to dissect the contribution of its coarse-to-fine autoregressive design. (cf. § 4.2) (4). Extended context: Can MSAR effectively benefit from longer look-back windows and better exploit historical dependencies? (cf. § 4.3) (5). Sensitivity analysis: How do the number of autoregressive layers and the choice of interval sizes affect forecasting accuracy? (cf. § 4.3)

Dataset. For the long-term forecasting experiments, we utilize a diverse set of datasets to evaluate the robustness and generalizability of our models across different domains. These datasets include ECL, ETT (4 subsets), Exchange, Traffic, Weather, Solar-Energy, and Wind.4 Additionally, for the short-term forecasting experiments, the PEMS (4 subsets) dataset is employed, specifically designed for short-term traffic prediction tasks. Following TSLib [33], we adopt the same dataset splits, data preprocessing steps and ensure no last batch is dropped during training. For long-term forecasting, we use prediction lengths of {96, 192, 336, 720}, while for short-term forecasting, we use a prediction length of 12.

Baselines. To evaluate the performance of our proposed MSAR framework, we compare it against several state-of-the-art (SOTA) methods, which include SimpleTM [23], DUET [29], iTransformer [12], PatchTST [11], and DLinear [8].

Unified experiment settings. To fairly evaluate the effectiveness of MSAR, we designed two experimental setups: (1). Baseline vs. Baseline + MSAR: In this experiment, we compare the performance of the baseline models with and without MSAR, ensuring that all hyperparameters remain consistent across both setups. (2). Full Hyperparameter Search: For this experiment, we perform a comprehensive search over the full hyperparameter space to ensure that all models are evaluated with the same range of parameter values, allowing for a fair comparison. Detailed experimental settings can be found in the Appendix § 7.3.

Forecasting Results

Performance Study. Table 2 reports average results across four horizons 96, 192, 336, 720 with all methods extracting features from a look-back window of length 336. Across ten datasets and five backbones, MSAR improves most model–dataset pairs. For instance, DLinear’s MSE on Exchange drops from 0.524 to 0.372, and PatchTST’s MAE on Weather decreases from 0.256 to 0.253. These consistent gains across both lightweight linear models and Transformer-style architectures (PatchTST, iTransformer, DUET, SimpleTM) highlight the generality of MSAR as a plug-in framework. Full per-horizon results are given in Appendix § 8.

DatasetMetricSimpleTM BaseSimpleTM +MSARDUET BaseDUET +MSARiTransformer BaseiTransformer +MSARPatchTST BasePatchTST +MSARDLinear BaseDLinear +MSAR
ECLMSE0.1640.1680.1780.1770.1670.1600.1680.1660.1700.168
MAE0.2540.2590.2770.2760.2570.2520.2580.2540.2690.259
ETTh1MSE0.4240.4100.4030.4040.4750.4490.4280.4130.4200.419
MAE0.4370.4250.4150.4160.4590.4550.4390.4280.4300.426
ETTh2MSE0.3770.3510.3550.3450.4050.3950.4040.3480.3970.391
MAE0.4010.3880.3900.3840.4200.4190.4220.3840.4140.412
ETTm1MSE0.3540.3450.3520.3510.3880.3590.3710.3450.3540.353
MAE0.3750.3710.3700.3690.3940.3840.3790.3690.3700.369
ETTm2MSE0.2830.2560.2550.2550.2810.2730.2580.2560.2600.259
MAE0.3220.3080.3090.3080.3350.3280.3140.3090.3140.313
ExchangeMSE0.4380.3520.3940.3720.4030.4370.4650.3640.5240.372
MAE0.4330.4010.4230.4080.4440.4620.4620.4080.4550.407
TrafficMSE0.4320.4330.4440.4450.4270.4190.4260.4180.4470.452
MAE0.2950.2940.3030.3020.2920.2820.2760.2670.3020.278
WeatherMSE0.2270.2260.2480.2460.2440.2270.2330.2300.2460.245
MAE0.2540.2530.2720.2710.2700.2620.2560.2530.2780.275
Solar-EnergyMSE0.2700.2420.2600.2500.2160.1960.2130.2060.2550.269
MAE0.2920.2670.2450.2410.2270.2250.2320.2290.3150.236
WindMSE0.7830.7690.7680.7670.7800.7320.7760.7610.7490.750
MAE0.6870.6770.6750.6750.6890.6640.6820.6750.6770.677

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 § 8.

We additionally compare methods under a shared hyperparameter search space; the results and analysis are provided in Appendix § 8.

Stratified Performance Analysis. To further contextualize the magnitude and consistency of improvements, we conduct a stratified analysis using dataset meta-features reported in TFB [34] benchmark, including seasonality strength, trend/non-stationarity, noise ratio, input dimensionality, and effective horizon difficulty. Three observations emerge.

❶ Larger gains on datasets with strong periodicity and stable seasonal structure. ETTh1/ETTh2 and ETTm1/ETTm2 are categorized as highly periodic and low-to-moderate noise. MSAR achieves the largest improvements here across nearly all backbones. Coarse scales reliably capture long-range seasonal cycles, providing stable guidance for finer-scale refinement.

❷ MSAR is particularly effective for long horizons where long-range structure dominates. The strongest improvements appear at horizons 336 and 720 on the ETT series and Wind datasets. This is consistent with the coarse-to-fine design: coarse-scale forecasting models slow-varying patterns, while finer scales reconstruct higher-frequency details.

❸ Moderate but consistent gains on high-noise or weakly periodic datasets. Traffic—characterized by high stochasticity and weak periodicity—shows smaller but still stable improvements. In these settings, coarse-scale signals carry limited structure, and irregular fine-scale fluctuations dominate; MSAR nevertheless improves long-horizon accuracy by stabilizing coarse-level guidance.

Training Efficiency. Beyond forecasting accuracy, we further evaluate the efficiency of MSAR. Figure 3 plots training time (x-axis), error (y-axis), and GPU memory (bubble size) for five representative baselines and their MSAR-augmented counterparts on Weather and ETTm1 (H=336H=336, T=192T=192). We report two variants of MSAR: full and flexible.

Efficiency–accuracy trade-off plots for Weather and ETTm1

Figure 3. Efficiency–accuracy trade-off on Weather (21 variates) and ETTm1 (7 variates). Different symbols indicate model variants: circles (●) represent the baseline, crosses (✖) represent the full version of MSAR, and triangles (▲) represent the flexible version of MSAR.

The full variant applies a unified look-back length of 336 for both forecasting and imputation. In this case, the forecasting module consumes a compressed 56-length input (via 6×6 \times downsampling), while the imputation module operates on the full 336 tokens. The flexible variant instead shortens the imputation look-back to 192, exploiting MSAR’s plug-in, model-agnostic design. This simple change leads to a substantial reduction in training cost while often improving accuracy.

MSAR incurs additional training-time overhead. Although each iteration is lighter (lower ms/iter) due to scale decoupling, the use of a two-stage training strategy to mitigate teacher-forcing-induced train–test discrepancy leads to about 1.67× longer wall-clock time than single-stage baselines. This cost is offset by more stable convergence, improved inference behavior, and reduced memory usage from decoupled per-scale forecasters, yielding a favorable efficiency–accuracy trade-off across diverse backbones.

Inference Efficiency. We further evaluate inference-time efficiency in Appendix § 9. MSAR reduces inference latency for PatchTST [11], DLinear [8], and TimesNet [5]—often by nearly a factor of two—because scale-wise forecasting shortens the effective sequence length processed at the fine scale. For iTransformer [12], MSAR introduces only a slight overhead, which is expected given its variate-wise embedding over the full sequence, where reducing token count does not yield proportional savings.

Ablation Study

In this section, we ask: where does the performance gain of MSAR come from? We ablate three design components: ① Multiscale, ② Alignment modeling in encoder, and ③ Scale-wise Autoregression (AR), using PatchTST [11] as the backbone. Alignment enforces cross-scale consistency between the input XX and output YY.

Table 3 shows that no single component alone is sufficient. Using only ① outperforms ② and ③, indicating the benefit of multi-resolution context, but its gains are limited by scale mixing. ② behaves similarly to a single-scale baseline, while ③ yields the weakest performance due to the lack of proper input–output alignment. Pairwise combinations improve performance, with ①+② consistently outperforming ①+③, highlighting the importance of alignment in mitigating multiscale interference.

SettingECLETTm1TrafficWeather
MSEMAEMSEMAEMSEMAEMSEMAE
①0.1990.2900.3830.3950.4720.3060.2550.277
②0.2050.2920.3870.3980.4860.3140.2640.283
③0.2120.2980.3950.4050.4950.3220.2710.289
①+②0.1840.2770.3790.3890.4560.2930.2440.269
①+③0.1910.2830.3810.3930.4600.2970.2470.272
①+②+③0.1720.2630.3550.3740.4410.2800.2340.253
①+②+③*0.1810.2730.3750.3860.4510.2900.2410.267
①+②+③†0.1820.2760.3660.3830.4440.2870.2370.264

Table 3. Ablation study on MSAR components over four datasets. ① = Multiscale, ② = Alignment, ③ = Scale-wise AR. Results with *: no joint training; †: no full pred-len imputation loss.

When all components are combined (①+②+③), MSAR achieves the best results across all datasets. The full design removes encoder-side scale mixing, preserves strict input–output alignment, and enables clean scale-wise autoregressive refinement. The consistent gains over all ablated variants confirm that MSAR’s improvements stem from the synergistic interaction of all three components rather than any single one.

Joint training and full-horizon supervision. The last two rows of Table 3 isolate the training choices from the architectural components. On ETTm1, removing joint training increases MSE from 0.355 to 0.375, while removing the full prediction-window imputation loss increases it to 0.366. Both changes also degrade MSE and MAE on ECL, Traffic, and Weather. These comparisons support two complementary roles: joint training exposes the refinement modules to predicted coarse-scale context, while full-horizon supervision allows fine-scale modules to correct inherited predictions as well as fill missing positions. The gains therefore depend on inference-aligned conditioning and supervision, not merely on adding scales or layers.

Model Analysis

Extended context. We vary the look-back window T∈{96,192,336,720,1152}T \in\{96, 192, 336, 720, 1152\} on ETTm1, Weather, and Electricity under two prediction horizons (H=96H = 96 and H=720H = 720). Figure 4 compares MSAR with five baselines. Longer histories generally reduce MSAR’s error, with substantial gains when extending the context from 96 to 336 steps. At T=1152T = 1152, MSAR remains among the strongest methods in all six dataset–horizon settings, showing that the framework can benefit from historical information beyond the fixed input length used in Table 2.

Impact of look-back window length on forecasting performance

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=96H = 96 and H=720H = 720). Each curve shows the error as the input sequence length increases from 96 to 1152.

The benefit is not uniformly monotonic. On ETTm1 with H=96H = 96, MSAR’s error increases from T=720T = 720 to 1152, whereas its long-horizon error continues to decrease. Electricity shows diminishing gains after the initial context extension, and several baselines fluctuate on ETTm1 and Weather. Thus, a longer history is useful only when the model can extract relevant structure; context length should still be selected for each dataset and horizon.

Sensitivity Analysis. We further investigate the sensitivity of MSAR with respect to the choice of interval list and the number of layers on the ETTh1 dataset. As shown in Figure 5(a), the top row reports the average performance of five backbones under different interval lists, with MSE and MAE shown separately. The results remain stable across varying downsampling factors, indicating that MSAR is largely insensitive to the specific interval choices and thus requires minimal tuning. Figure 5(b) illustrates the impact of the number of layers. While adding one or two imputation layers consistently improves both MSE and MAE over the baseline, further increasing the number of layers yields diminishing returns and in some cases even slight degradation. These observations confirm that MSAR achieves robust performance across settings, with a favorable trade-off between accuracy and complexity.

MSAR sensitivity plots on ETTh1

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.

Full-horizon refinement. Full-horizon supervision allows fine-scale modules to revise inherited predictions rather than treating them as fixed outputs. Coarse predictions serve as conditioning inputs, while each finer module predicts the entire target window using its higher-resolution history. Joint training exposes this correction process to predicted context instead of exact target values. Together, these choices provide a mechanism for limiting inherited errors, supported by the ablations above, without guaranteeing that errors disappear as more scales are added.

Interaction with channel modeling. MSAR applies the same temporal stride to every channel, retaining all CC variables at each selected time point. Scale construction neither partitions nor mixes variables, so each backbone retains its own channel-modeling mechanism. Channel-independent backbones such as PatchTST and DLinear remain channel independent, while iTransformer and DUET can model cross-channel relationships within each scale. This distinction concerns the architecture: subsampling changes the temporal information available at coarse scales and need not preserve every cross-channel dependency. Finer modules also receive their own higher-resolution historical inputs, allowing them to use information absent from coarse predictions. MSAR therefore complements the backbone’s variable modeling through temporal scale alignment.

Limitations and Future Work. MSAR can propagate coarse-scale errors, adds training stages, and does not improve every backbone–dataset setting. The reported improvements are based on point estimates, and we do not systematically characterize variability across random seeds. The flexible variant reduces computation, but real-time deployment may require adaptive scheduling or scale selection. Future work includes joint multiscale pretraining and scaling studies, residual corrections to inherited predictions, and refinement under structured noise [35] to improve robustness to imperfect coarse context.

Conclusions

In this study, we propose the Modular Scale-wise Autoregressive Framework (MSAR) to overcome the limitations of existing time series forecasting methods that suffer from scale misalignment and limited long-context utilization. Unlike prior approaches that apply multiscale modeling only to the input sequence or rely on single-scale decoding, MSAR performs fully scale-aligned modeling across both inputs and outputs, enabling each scale to be modeled independently. By introducing a scale-wise autoregressive strategy, MSAR progressively refines predictions from coarse to fine resolutions through an efficient, modular information flow. Our work highlights the importance of aligning temporal granularity in both modeling and generation, paving the way for scalable and robust multiscale forecasting.

Acknowledgments and Disclosure of Funding

Funding. This work was supported by the Innovation Funding of ICT, CAS under Grant No. E661110000. Additional support was provided by the Key Scientific Research Project of Higher Education Institutions of Henan Province (No. 24A520006) and the Research Fund of the School of Computer and Information Engineering (No. K24061Y), both awarded to Nana Wang.

Competing Interests. The authors declare no competing interests.

Implementation Details

Dataset Descriptions

We conduct experiments on fourteen real-world multivariate time series datasets, including ten long-term forecasting datasets and four short-term forecasting datasets. The datasets span diverse domains including electricity, weather, traffic, and renewable energy, providing a range of settings for evaluating our approach. Table 4 summarizes the statistics of all datasets. Specifically:

DatasetDimDataset Size (Train/Val/Test)Frequency
ETTh1, ETTh27(8545, 2881, 2881)Hourly
ETTm1, ETTm27(34465, 11521, 11521)15 min
Weather21(36792, 5271, 10540)10 min
Traffic862(12185, 1757, 3509)Hourly
Electricity321(18317, 2633, 5261)Hourly
Exchange-rate8(5120, 665, 1422)Daily
Solar-energy137(36792, 5271, 10540)10 min
Wind9(30468, 4286, 8665)Hourly
PEMS03358(15617, 5135, 5135)5 min
PEMS04307(10172, 3375, 3375)5 min
PEMS07883(16911, 5622, 5622)5 min
PEMS08170(10690, 3548, 3548)5 min

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].

  • Electricity Transformer Temperature (ETT) [36]: Four subsets are provided. ETTh1/ETTh2 are hourly, while ETTm1/ETTm2 are recorded every 15 minutes. All series are collected from two transformers.

  • Weather [37]: 21 meteorological indicators collected every 10 minutes in Germany during 2020 from the Max Planck Biogeochemistry Institute.

  • Traffic [37]: Hourly road occupancy rates from 862 sensors on San Francisco Bay Area freeways (2015–2016).

  • Electricity [37]: Hourly consumption of 321 clients from 2012–2014.

  • Exchange-rate [37]: Daily exchange rates of 8 countries from 1990–2016.

  • Solar-energy [38]: 10-minute solar power production from 137 photovoltaic plants in 2006.

  • Wind5: Hourly wind speed and power generation from four real-world wind farms.

  • PEMS: High-frequency traffic measurements with 5-minute resolution. We use four subsets (PEMS03, PEMS04, PEMS07, and PEMS08), each corresponding to different spatial sensor networks.

Evaluation Metrics

We follow standard practices in time series forecasting and adopt different metrics for short-term and long-term datasets.

Short-term forecasting. For short-term benchmarks such as PEMS, we evaluate models using three complementary error measures:

  • Mean Absolute Error (MAE): 1T∑t=1T∣y^t−yt∣\frac{1}{T}\sum_{t=1}^{T}|\hat{y}_t-y_t|

  • Mean Absolute Percentage Error (MAPE): 1T∑t=1T∣y^t−yt∣∣yt∣\frac{1}{T}\sum_{t=1}^{T}\frac{|\hat{y}_t-y_t|}{|y_t|}

  • Root Mean Squared Error (RMSE): 1T∑t=1T(y^t−yt)2\sqrt{\frac{1}{T}\sum_{t=1}^{T}(\hat{y}_t-y_t)^2}

MAE reflects absolute deviation, MAPE measures relative deviation with respect to ground truth magnitude, and RMSE penalizes large errors more heavily. Together, these metrics provide a comprehensive evaluation of accuracy and robustness.

Long-term forecasting. For long-term benchmarks such as ETT, Weather, and Exchange, we report Mean Squared Error (MSE) and Mean Absolute Error (MAE):

  • Mean Squared Error (MSE): 1T∑t=1T(y^t−yt)2\frac{1}{T}\sum_{t=1}^{T}(\hat{y}_t-y_t)^2

  • Mean Absolute Error (MAE): 1T∑t=1T∣y^t−yt∣\frac{1}{T}\sum_{t=1}^{T}|\hat{y}_t-y_t|

These metrics are widely adopted in prior work [11], and capture both variance-sensitive and absolute error perspectives. MSE emphasizes stability for long horizons by penalizing large deviations, while MAE provides scale-consistent interpretability. All metrics are averaged across all variates and prediction horizons. Lower values indicate better forecasting performance.

Experiment details

To fairly evaluate the effectiveness of MSAR, we design two experimental setups: (1) Baseline vs. Baseline+MSAR: we compare the performance of each baseline model with and without MSAR, while keeping all hyperparameters consistent across both settings. (2) Full Hyperparameter Search: we perform a comprehensive grid search over the full hyperparameter space so that all models are tuned under the same parameter search range.

Settings for (1). For the baseline vs. baseline+MSAR setup, we adopt a unified input length (seq_len = 336) across all datasets, with the detailed hyperparameters summarized in Table 5. For MSAR, we fix the interval list to interval_list = [6, 1], which corresponds to seqlen_list = [56, 336] for coarse- and fine-scale autoregression. This configuration guarantees that the baseline models and their MSAR-augmented counterparts are evaluated under identical input conditions, thereby ensuring a fair comparison.

Dataset / Configurationsdmodeld_{model}dffd_{ff}LayersHeadsBatch SizeEpochsLR*Patience
ECL12825612161010−310^{-3}3
Traffic12825612161010−310^{-3}3
Solar12825612161010−310^{-3}3
ETTh1512204828321010−310^{-3}3
ETTh2512204828321010−310^{-3}3
ETTm1512204828321010−310^{-3}3
ETTm2512204828321010−310^{-3}3
Weather512204828321010−310^{-3}3
Exchange512204828321010−310^{-3}3
Wind512204828321010−310^{-3}3

Table 5. Experiment configuration of MSAR in Table 6. All experiments use the Adam optimizer with a learning rate of 10−310^{-3} and early stopping (patience = 3). * LR means the initial learning rate.

Setup for (2). For the full hyperparameter search, all methods are tuned under an identical search space to ensure fairness. Specifically, the candidate ranges are:

  • seq_len ∈{96,192,320,512}\in\{96, 192, 320, 512\},

  • d_model ∈{16,128,512}\in\{16, 128, 512\},

  • e_layers ∈{1,3,5}\in\{1, 3, 5\},

  • epoch ∈{50}\in\{50\},

  • learning rate ∈{1×10−5,1×10−4,1×10−3,1×10−2,0.05}\in\{1 \times10^{-5}, 1 \times10^{-4}, 1 \times10^{-3}, 1 \times10^{-2}, 0.05\}.

Other hyperparameters are fixed across all models, including batch_size = 32, n_heads = 8, and d_layers = 1.

This unified search space covers model width, depth, horizon length, and optimization settings, and is applied consistently across all baselines. Therefore, the reported results reflect the intrinsic modeling capability of each method rather than differences in hyperparameter tuning.

All experiments are implemented in Pytorch [39], and conducted on two machines, each equipped with four NVIDIA 4090 GPUs.

More Results

Full Results in Table 2. Table 6 reports the full long-term forecasting results across all prediction lengths {96,192,336,720}\{96, 192, 336, 720\} on the benchmark datasets. Here we provide the complete results to ensure full transparency and allow a more fine-grained comparison across different forecasting lengths.

ModelsMetricSimpleTM Base MSESimpleTM Base MAESimpleTM +Ours MSESimpleTM +Ours MAEDUET Base MSEDUET Base MAEDUET +Ours MSEDUET +Ours MAEiTransformer Base MSEiTransformer Base MAEiTransformer +Ours MSEiTransformer +Ours MAEPatchTST Base MSEPatchTST Base MAEPatchTST +Ours MSEPatchTST +Ours MAEDLinear Base MSEDLinear Base MAEDLinear +Ours MSEDLinear +Ours MAE
Electricity 960.1340.2260.1370.2310.1490.2510.1470.2500.1350.2290.1320.2260.1380.2310.1360.2270.1430.2420.1410.234
Electricity 1920.1540.2450.1520.2440.1630.2640.1620.2630.1530.2450.1490.2410.1540.2460.1520.2410.1570.2550.1550.246
Electricity 3360.1750.2680.1670.2580.1810.2810.1790.2790.1690.2610.1640.2570.1710.2610.1680.2570.1720.2720.1700.262
Electricity 7200.1910.2780.2140.3030.2200.3110.2190.3100.2090.2940.1960.2850.2100.2940.2070.2900.2070.3050.2040.293
Electricity Avg0.1640.2540.1680.2590.1780.2770.1770.2760.1670.2570.1600.2520.1680.2580.1660.2540.1700.2690.1680.259
ETTm1 960.2910.3370.2820.3300.2920.3350.2930.3350.3180.3550.3000.3440.3170.3430.2800.3270.2930.3330.2930.333
ETTm1 1920.3300.3630.3240.3570.3320.3590.3300.3560.3700.3830.3370.3700.3350.3610.3240.3560.3320.3550.3320.355
ETTm1 3360.3690.3840.3570.3810.3650.3780.3630.3760.4000.4030.3720.3930.3810.3840.3610.3790.3660.3760.3660.377
ETTm1 7200.4250.4160.4160.4150.4190.4090.4190.4090.4620.4340.4280.4270.4510.4270.4140.4120.4260.4170.4220.411
ETTm1 Avg0.3540.3750.3450.3710.3520.3700.3510.3690.3880.3940.3590.3840.3710.3790.3450.3690.3540.3700.3530.369
ETTm2 960.1770.2550.1650.2460.1630.2480.1630.2470.1790.2660.1720.2550.1730.2580.1640.2460.1650.2480.1650.247
ETTm2 1920.2510.3020.2210.2870.2190.2850.2190.2850.2540.3230.2280.2980.2190.2880.2240.2880.2220.2880.2220.288
ETTm2 3360.3050.3380.2720.3190.2730.3220.2720.3220.3000.3500.2960.3510.2790.3270.2790.3240.2760.3270.2760.326
ETTm2 7200.3990.3940.3640.3780.3650.3790.3640.3790.3890.3990.3970.4090.3610.3810.3550.3790.3750.3930.3710.391
ETTm2 Avg0.2830.3220.2560.3080.2550.3090.2550.3080.2810.3350.2730.3280.2580.3140.2560.3090.2600.3140.2590.313
ETTh1 960.3790.4010.3610.3890.3590.3830.3580.3810.4090.4150.3870.4060.3830.4030.3590.3860.3650.3830.3650.383
ETTh1 1920.4230.4300.4100.4180.3980.4060.3980.4060.5070.4620.4290.4360.4090.4240.3970.4120.4170.4260.4050.409
ETTh1 3360.4400.4430.4280.4310.4260.4220.4280.4230.4590.4510.4570.4570.4400.4460.4220.4330.4430.4360.4400.431
ETTh1 7200.4550.4750.4400.4600.4300.4500.4330.4520.5260.5060.5230.5190.4810.4820.4750.4790.4560.4730.4670.481
ETTh1 Avg0.4240.4370.4100.4250.4030.4150.4040.4160.4750.4590.4490.4550.4280.4390.4130.4280.4200.4300.4190.426
ETTh2 960.2950.3420.2770.3340.2730.3310.2720.3300.3430.3730.2920.3480.2920.3460.2880.3380.2780.3330.2780.333
ETTh2 1920.3610.3880.3470.3790.3470.3820.3380.3750.4100.4130.3700.3990.3540.3850.3500.3780.3500.3820.3520.385
ETTh2 3360.4260.4280.3720.4050.3940.4150.3710.4020.4190.4340.4210.4310.3830.4070.3650.4000.4030.4240.3890.413
ETTh2 7200.4270.4460.4080.4320.4050.4330.4000.4300.4480.4610.4970.4980.5870.5490.3870.4200.5560.5180.5460.515
ETTh2 Avg0.3770.4010.3510.3880.3550.3900.3450.3840.4050.4200.3950.4190.4040.4220.3480.3840.3970.4140.3910.412
Exchange 960.0940.2160.0900.2120.0920.2130.0880.2060.1090.2390.1070.2370.0980.2210.0900.2100.0820.2010.0950.214
Exchange 1920.1900.3130.1830.3080.1890.3070.1840.3020.2660.3770.2690.3670.2100.3290.1950.3170.1740.2990.1540.283
Exchange 3360.3310.4170.3420.4240.3920.4560.3540.4280.3960.4670.5790.5660.4680.5060.3790.4430.3470.4270.2760.392
Exchange 7201.1380.7840.7920.6590.9010.7140.8630.6970.8410.6910.7910.6781.0820.7930.7910.6611.4930.8930.9610.737
Exchange Avg0.4380.4330.3520.4010.3940.4230.3720.4080.4030.4440.4370.4620.4650.4620.3790.4080.5240.4550.3720.407
Traffic 960.4060.2850.3980.2750.4240.2880.4250.2900.3980.2750.3990.2670.4020.2640.3990.2570.4300.2690.4340.268
Traffic 1920.4210.2890.4250.2900.4330.2960.4340.3030.4170.2860.4070.2820.4290.2760.4160.2650.4350.3020.4440.277
Traffic 3360.4350.2960.4370.2960.4450.3040.4460.3030.4310.2940.4420.2820.4290.2760.4090.2650.4470.3080.4530.277
Traffic 7200.4650.3090.4720.3160.4740.3220.4740.3200.4630.3120.4560.3010.4560.2930.4470.2810.4760.3270.4770.293
Traffic Avg0.4320.2950.4330.2940.4440.3030.4450.3020.4270.2920.4190.2820.4260.2760.4180.2670.4470.3020.4520.278
Weather 960.1480.1880.1470.1870.1760.2140.1720.2110.1620.2040.1520.1910.1550.1890.1510.1860.1780.2150.1780.213
Weather 1920.1930.2320.1890.2270.2170.2510.2170.2510.2130.2490.1950.2380.1940.2300.1950.2300.2180.2520.2170.252
Weather 3360.2470.2740.2420.2700.2650.2880.2630.2860.2630.2870.2440.2780.2560.2790.2460.2710.2610.2910.2610.292
Weather 7200.3180.3230.3250.3270.3330.3340.3320.3340.3390.3400.3150.3380.3270.3270.3270.3260.3260.3540.3220.344
Weather Avg0.2270.2540.2260.2530.2480.2720.2460.2710.2440.2800.2270.2620.2330.2560.2300.2530.2460.2780.2450.275
Solar-Energy 960.2340.2630.2220.2640.2240.2220.2200.2220.1880.2100.1740.2070.2040.2290.1860.2170.2230.2930.2340.214
Solar-Energy 1920.2660.2870.2510.2780.2560.2420.2480.2390.2130.2260.2040.2330.2110.2250.2010.2240.2510.3110.2660.234
Solar-Energy 3360.2890.3040.2400.2630.2790.2570.2720.2540.2280.2350.2000.2330.2170.2360.2120.2380.2720.3270.2890.248
Solar-Energy 7200.2900.3150.2540.2640.2800.2600.2610.2500.2330.2370.2040.2270.2200.2360.2260.2370.2740.3300.2880.248
Solar-Energy Avg0.2700.2920.2420.2670.2600.2450.2500.2410.2160.2270.1960.2250.2130.2320.2060.2290.2550.3150.2690.236
Wind 960.7500.6600.7340.6510.7200.6440.7200.6440.7410.6640.7030.6400.7330.6510.7200.6460.7090.6460.7100.646
Wind 1920.7700.6810.7590.6690.7590.6690.7590.6690.7790.6850.7350.6630.7580.6770.7580.6710.7430.6720.7440.672
Wind 3360.7930.6950.7800.6830.7800.6830.7800.6830.7890.6960.7440.6720.8000.6920.7710.6810.7620.6850.7620.685
Wind 7200.8180.7110.8040.7030.8110.7040.8100.7040.8120.7120.7450.6790.8110.7060.7940.7000.7810.7030.7820.703
Wind Avg0.7830.6870.7690.6770.7680.6750.7670.6750.7800.6890.7320.6640.7760.6820.7610.6750.7490.6770.7500.677

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.

Full Hyperparameter Search. Table 7 compares MSAR with five baselines after tuning all methods within a shared search space. The look-back window is selected from {96,192,320,512}\{96, 192, 320, 512\}, with common candidate ranges for model width, depth, and learning rate; the full search protocol is provided in Appendix § 7.3. We report MSE and MAE at four prediction lengths on the four ETT datasets and Weather. MSAR achieves the lowest average MSE and MAE on ETTm1, ETTm2, and ETTh1. On ETTh2, it achieves the lowest average MAE, while DUET obtains a lower average MSE.

ModelsMSARSimpleTMDUETiTransformerPatchTSTDLinear
Ours[2025][2025][2024][2023][2023]
MetricMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
ETTm1960.2890.3350.2800.3370.3000.3450.2990.3480.2890.3430.2990.343
ETTm11920.3250.3550.3260.3650.3350.3660.3340.3730.3320.3670.3360.367
ETTm13360.3550.3760.3620.3870.3640.3830.3700.3940.3620.3890.3670.385
ETTm17200.4100.4060.4170.4220.4170.4130.4220.4260.4080.4200.4200.418
ETTm1Avg.0.3450.3680.3460.3780.3540.3770.3560.3850.3480.3800.3560.378
ETTm2960.1600.2460.1680.2520.1620.2530.1800.2630.1650.2570.1650.260
ETTm21920.2160.2840.2230.2970.2170.2900.2360.3070.2230.2940.2190.297
ETTm23360.2690.3210.2730.3280.2700.3250.2870.3400.2720.3330.2900.345
ETTm27200.3590.3770.3640.3860.3610.3840.3640.3910.3630.3860.3750.403
ETTm2Avg.0.2510.3070.2570.3160.2530.3130.2670.3250.2560.3180.2620.326
ETTh1960.3560.3810.3750.3940.3640.3920.3800.4010.3700.3970.3680.393
ETTh11920.3950.4070.4120.4290.3970.4130.4230.4310.4040.4230.4000.417
ETTh13360.4260.4240.4290.4430.4220.4320.4370.4500.4230.4370.4300.441
ETTh17200.4280.4490.4470.4720.4430.4630.4600.4710.4440.4640.4770.497
ETTh1Avg.0.4010.4170.4160.4340.4060.4250.4250.4380.4100.4300.4190.437
ETTh2960.2670.3280.2910.3390.2650.3340.2940.3450.2900.3420.2860.351
ETTh21920.3310.3620.3490.3900.3240.3730.3580.3950.3540.3970.3520.394
ETTh23360.3640.3990.3810.4130.3520.3990.3850.4150.3840.4130.4390.456
ETTh27200.3970.4300.4100.4420.3930.4320.4120.4360.4080.4420.5700.530
ETTh2Avg.0.3400.3800.3580.3960.3340.3840.3620.3980.3590.3990.4120.433
Weather960.1450.1890.1430.1950.1680.2210.1550.2060.1450.1940.1700.229
Weather1920.1880.2300.1880.2360.2120.2580.1990.2490.1910.2370.2120.268
Weather3360.2570.2840.2380.2760.2580.2920.2470.2840.2430.2790.2580.307
Weather7200.3230.3320.3120.3270.3240.3380.3180.3360.3130.3300.3210.359
WeatherAvg.0.2280.2590.2200.2590.2400.2770.2300.2690.2230.2600.2400.291
1st Count3766010

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 § 7.3.

On Weather, MSAR ties SimpleTM in average MAE at the reported precision, while SimpleTM achieves a lower average MSE. These results show that MSAR remains competitive when the baselines are tuned beyond the fixed-setting comparison.

Short-term forecasting results.

Table Table 8 reports short-term forecasting results on the PEMS datasets with a prediction length of 12. We evaluate three common metrics—MAE, MAPE, and RMSE—to provide a comprehensive view of performance. MSAR lowers MAE for SimpleTM, iTransformer, PatchTST, and DLinear on all four subsets. DUET benefits less consistently, with small MAE and RMSE increases on PEMS04 and PEMS08. Thus, the gains extend to short-term forecasting, but their size and consistency depend on the backbone.

DatasetMetricSimpleTM BaseSimpleTM +OursDUET BaseDUET +OursiTransformer BaseiTransformer +OursPatchTST BasePatchTST +OursDLinear BaseDLinear +Ours
PEMS03MAE19.14715.62718.85318.40215.87015.30816.97415.20917.41116.743
PEMS03MAPE19.52715.79622.03120.69815.94515.45919.37615.90515.87615.509
PEMS03RMSE30.30724.93329.60729.24025.56324.64826.51724.31029.61927.909
PEMS04MAE27.19221.65124.58124.69322.03320.61723.00521.02224.35922.546
PEMS04MAPE18.06413.46218.65918.20214.03713.01915.77913.84614.61613.538
PEMS04RMSE41.71934.42537.48637.61434.83433.02335.61633.30038.96636.244
PEMS07MAE26.99922.77026.86226.57723.06421.86324.01021.91125.51924.655
PEMS07MAPE12.1589.67513.40113.25810.1199.54410.8219.31110.86910.384
PEMS07RMSE41.25735.83340.31440.11436.33034.76236.71133.9240.31339.109
PEMS08MAE22.26116.41219.92219.97617.35015.85118.59116.05720.35318.090
PEMS08MAPE14.12410.08613.63113.45910.9639.99012.56910.41912.37611.047
PEMS08RMSE34.45526.05930.19930.72027.45925.31228.17225.26732.65829.043

Table 8. Short-term forecasting results on PEMS datasets (prediction length = 12). Lower MAE, MAPE, and RMSE indicate better performance.

Efficiency Study

To assess the inference efficiency of our MSAR framework, we report the wall-clock runtime of four backbone architectures—DLinear [8], TimesNet [5], PatchTST [11], and iTransformer [12]—both with and without MSAR integration. Additionally, we compare against SegRNN [4], a representative autoregressive baseline that sequentially generates future time steps.

Figure 6 compares the inference time of baseline models with their MSAR-enhanced counterparts under identical input-output settings (input length = 960, prediction length = 336, batch size = 32). For most backbones, including TimesNet [5] and PatchTST [11], the hierarchical design of MSAR improves both prediction accuracy and runtime by reducing redundant computation through coarse-scale anchoring. However, we observe an increase in inference time when integrating MSAR with iTransformer [12]. This is expected, as iTransformer encodes inputs via a global linear projection into the latent space without explicit dependency on sequence length. The hierarchical multi-scale processing in MSAR introduces additional overhead in such architectures, where lookback windows do not otherwise impact runtime.

Inference efficiency of baseline models vs. MSAR-enhanced models

Figure 6. Inference efficiency of baseline models vs. MSAR-enhanced models (batch size 32, input=960, pred=336).

Table 9 collectively demonstrates the inference efficiency of MSAR-enhanced models compared to autoregressive baselines on the ETTh1 dataset. It shows that MSAR consistently achieves significant inference speedups over SegRNN [4], as it avoids full-step autoregression by leveraging parallel decoding within each scale. MSAR achieves favorable efficiency-accuracy trade-offs across backbones, demonstrating its generalizability and practical utility for scalable time series forecasting.

ModelHInference Time (s)
SegRNN⁶3367.23
SegRNN¹²3364.06
SegRNN⁶7208.53
SegRNN¹²7205.01
PatchTST*3362.13
PatchTST*7202.30
iTransformer*3361.84
iTransformer*7201.94
DLinear*3360.31
DLinear*7200.36

Table 9. Inference time (in seconds) for different models and prediction lengths H. * indicates MSAR-enhanced models. SegRNN⁶ and SegRNN¹² denote segment length.

Additional Prediction Results

In this section, we provide additional qualitative results on the ETTh1 dataset. All experiments are conducted with a single-layer encoder (Layer=1), two attention heads (Head=2), and an interval configuration of interval_list = [3, 1]. We evaluate the model under four input sequence length settings: [32, 96], [32, 192], [32, 384], and [32, 512]. The visualizations below are generated using our proposed DUET+MSAR model, showing its prediction accuracy compared to the ground-truth time series.

The visualizations in Figure 7 are generated using our proposed DUET+MSAR model, illustrating its prediction accuracy compared to the ground-truth time series.

Visualization of model predictions and ground truth on the ETTh1 dataset

Figure 7. Visualization of model predictions and ground truth on the ETTh1 dataset. Results are produced by the DUET+MSAR model.

NeurIPS Paper Checklist

  1. Claims

Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope?

Answer: [Yes]

Justification: The introduction clearly states the three key contributions of MSAR: scale-wise aligned modeling, scale-wise autoregression, and a modular model-agnostic design. These claims are directly supported by the experimental results in § 4.1 and the ablation study in § 4.2.

Guidelines:

  • The answer [N/A] means that the abstract and introduction do not include the claims made in the paper.

  • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers.

  • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.

  • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.

  1. Limitations

Question: Does the paper discuss the limitations of the work performed by the authors?

Answer: [Yes]

Justification: The “Limitations and Future Work” paragraph at the end of the Experiments section discusses coarse-scale error propagation, additional training stages, and the lack of improvements in some backbone–dataset settings. It also acknowledges that variability across random seeds is not systematically characterized and that real-time deployment may require adaptive scheduling or scale selection.

Guidelines:

  • The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper.

  • The authors are encouraged to create a separate “Limitations” section in their paper.

  • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.

  • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.

  • The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon.

  • The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.

  • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.

  • While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.

Theory assumptions and proofs

Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?

Answer: [N/A]

Justification: The paper presents an algorithmic framework and empirical analysis without formal theoretical results requiring mathematical proofs.

Guidelines:

  • The answer [N/A] means that the paper does not include theoretical results.

  • All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.

  • All assumptions should be clearly stated or referenced in the statement of any theorems.

  • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.

  • Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.

  • Theorems and Lemmas that the proof relies upon should be properly referenced.

Experimental result reproducibility

Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?

Answer: [Yes]

Justification: The paper provides full experimental details in Appendix § 7.3, including dataset splits, hyperparameter search spaces, optimizer (Adam, lr=10−310^{-3}), training settings, and the MSAR-specific configuration (interval_list=[6, 1]). All experiments are implemented in PyTorch on NVIDIA 4090 GPUs.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.

  • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.

  • Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.

  • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset).

(d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.

Open access to data and code

Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?

Answer: [No]

Justification: The implementation is available at https://github.com/Oryxo/MSAR. The experiments use publicly available benchmark datasets, with descriptions and source references provided in Appendix § 7.§ 1. The repository includes a dependency list, data-format guidance, example training commands, configurable hyperparameters, and evaluation code. However, the current release does not provide complete data-preparation procedures and experiment-specific configurations for all reported results, so it does not yet satisfy the full reproducibility requirement of this question.

  • The answer [N/A] means that paper does not include experiments requiring code.

  • Please see the NeurIPS code and data submission guidelines (https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.

  • While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).

  • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.

  • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.

  • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.

  • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).

  • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.

Experimental setting/details

Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results?

Answer: [Yes]

Justification: Training and evaluation details are described in the Experiments section and Appendix § 7.3. This includes dataset splits following TSLib, the Adam optimizer with lr=10−3lr=10^{-3} and early stopping (patience=3), unified seq_len=336, and full hyperparameter search ranges.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.

  • The full details can be provided either with the code, in appendix, or as supplemental material.

Experiment statistical significance

Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [No]

Justification: The paper does not report error bars, confidence intervals, or statistical significance tests. The reported improvements are based on point estimates, and we do not systematically characterize variability across random seeds. Averaging over prediction lengths does not quantify seed-level uncertainty.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.

  • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).

  • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)

  • The assumptions made should be given (e.g., Normally distributed errors).

  • It should be clear whether the error bar is the standard deviation or the standard error of the mean.

  • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified.

  • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates).

  • If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.

Experiments compute resources

Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?

Answer: [Yes]

Justification: The appendix (Appendix § 7.3) states that all experiments are conducted on two machines each equipped with four NVIDIA 4090 GPUs. Training efficiency (wall-clock time and GPU memory) is further analyzed in Figure 3.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.

  • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.

  • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper).

Code of ethics

Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?

Answer: [Yes]

Justification: This work proposes a time series forecasting framework evaluated on publicly available benchmark datasets. It does not involve human subjects, sensitive data, or any application with foreseeable ethical risks.

Guidelines:

  • The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics.

  • If the authors answer [No], they should explain the special circumstances that require a deviation from the Code of Ethics.

  • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).

Broader impacts

Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?

Answer: [N/A]

Justification: This is a foundational research paper on time series forecasting methodology. There is no direct path to negative societal impacts from improving forecasting accuracy on standard benchmarks.

Guidelines:

  • The answer [N/A] means that there is no societal impact of the work performed.

  • If the authors answer [N/A] or [No], they should explain why their work has no societal impact or why the paper does not address societal impact.

  • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.

  • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.

  • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.

  • If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).

Safeguards

Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)?

Answer: [N/A]

Justification: The paper presents a time series forecasting framework without releasing pre-trained models or scraped datasets. There is no high risk for misuse associated with this work.

Guidelines:

  • The answer [N/A] means that the paper poses no such risks.

  • Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.

  • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.

  • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.

12. Licenses for existing assets

Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?

Answer: [Yes]

Justification: All datasets and baseline models used in the paper are properly cited in the references. The datasets (ETT, ECL, Weather, Traffic, PEMS, etc.) are standard public benchmarks widely used in the time series forecasting literature.

Guidelines:

  • The answer [N/A] means that the paper does not use existing assets.

  • The authors should cite the original paper that produced the code package or dataset.

  • The authors should state which version of the asset is used and, if possible, include a URL.

  • The name of the license (e.g., CC-BY 4.0) should be included for each asset.

  • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.

  • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.

  • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.

  • If this information is not available online, the authors are encouraged to reach out to the asset’s creators.

13. New assets

Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?

Answer: [N/A]

Justification: The paper does not introduce new datasets or models released as assets. MSAR is a plug-in framework applied on top of existing models and evaluated on existing datasets.

Guidelines:

  • The answer [N/A] means that the paper does not release new assets.

  • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.

  • The paper should discuss whether and how consent was obtained from people whose asset is used.

  • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.

14. Crowdsourcing and research with human subjects

Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?

Answer: [N/A]

Justification: The paper does not involve crowdsourcing or research with human subjects. All experiments are conducted on publicly available time series datasets.

Guidelines:

  • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.

  • Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.

  • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.

  1. Institutional review board (IRB) approvals or equivalent for research with human subjects

Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?

Answer: [N/A]

Justification: The paper does not involve human subjects or crowdsourcing; hence IRB approval is not required.

Guidelines:

  • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.

  • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.

  • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.

  • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

  1. Declaration of LLM usage

Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required.

Answer: [N/A]

Justification: LLMs are not used as part of the core methodology of this research. MSAR is a scale-wise autoregressive framework for time series forecasting that does not rely on large language models.

Guidelines:

  • The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.

  • Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.

References

  1. [1]David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International journal of forecasting, 2020.arxiv.org/abs/1704.04110
  2. [2]Shane Bergsma, Tim Zeyl, Javad Rahimipour Anaraki, and Lei Guo. C2FAR: Coarse-to-fine autoregressive networks for precise probabilistic forecasting. In NeurIPS, 2022.arxiv.org/abs/2312.15002
  3. [3]Shane Bergsma, Tim Zeyl, and Lei Guo. Sutranets: Sub-series autoregressive networks for long-sequence, probabilistic forecasting. In NeurIPS, 2023.arxiv.org/abs/2312.14880
  4. [4]Shengsheng Lin, Weiwei Lin, Wentai Wu, Feiyu Zhao, Ruichao Mo, and Haotong Zhang. Segrnn: Segment recurrent neural network for long-term time series forecasting. arXiv preprint arXiv:2308.11200, 2023.arxiv.org/abs/2308.11200
  5. [5]Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. In International Conference on Learning Representations, 2023.arxiv.org/abs/2210.02186
  6. [6]Huiqiang Wang, Jian Peng, Feihu Huang, Jince Wang, Junhui Chen, and Yifei Xiao. Micn: Multi-scale local and global context modeling for long-term series forecasting. In ICLR, 2023.
  7. [7]Tianping Zhang, Yizhuo Zhang, Wei Cao, Jiang Bian, Xiaohan Yi, Shun Zheng, and Jian Li. Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp structures. arXiv preprint arXiv:2207.01186, 2022.arxiv.org/abs/2207.01186
  8. [8]Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? In AAAI, 2023.arxiv.org/abs/2205.13504
  9. [9]Guoqi Yu, Jing Zou, Xiaowei Hu, Angelica I Aviles-Rivero, Jing Qin, and Shujun Wang. Revitalizing multivariate time series forecasting: Learnable decomposition with inter-series dependencies and intra-series variations modeling. In ICML, 2024.arxiv.org/abs/2402.12694
  10. [10]Peiwang Tang and Weitai Zhang. Unlocking the power of patch: Patch-based mlp for long-term time series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 12640–12648, 2025.arxiv.org/abs/2405.13575
  11. [11]Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. In The Eleventh International Conference on Learning Representations, 2023.arxiv.org/abs/2211.14730
  12. [12]Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024.arxiv.org/abs/2310.06625
  13. [13]Lorenzo Brigato, Rafael Morand, Knut Strømmen, Maria Panagiotou, Markus Schmidt, and Stavroula Mougiakakou. Position: There are no champions in long-term time series forecasting. arXiv preprint arXiv:2502.14045, 2025.arxiv.org/abs/2502.14045
  14. [14]Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y. Zhang, and JUN ZHOU. Timemixer: Decomposable multiscale mixing for time series forecasting. In ICLR, 2024.arxiv.org/abs/2405.14616
  15. [15]Zhipeng Liu, Peibo Duan, Xuan Tang, Baixin Li, Yongsheng Huang, Mingyang Geng, Changsheng Zhang, Bin Zhang, and Binwu Wang. Timeformer: Transformer with attention modulation empowered by temporal characteristics for time series forecasting. Expert Syst. Appl., 307(C), June 2026. ISSN 0957-4174. doi: 10.1016/j.eswa.2025.131040. URL https://doi.org/10.1016/j.eswa.2025.131040.
  16. [16]Amir B Geva. Scalenet-multiscale neural-network architecture for time series prediction. IEEE Transactions on neural networks, 1998.
  17. [17]Qiang Guo, Lexin Fang, Ren Wang, and Caiming Zhang. Multivariate time series forecasting using multiscale recurrent networks with scale attention and cross-scale guidance. TNNLS, 2023.DOI
  18. [18]Yunhao Zhang and Junchi Yan. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. In The eleventh international conference on learning representations, 2023.
  19. [19]Amin Shabani, Amir Abdi, Lili Meng, and Tristan Sylvain. Scaleformer: Iterative multi-scale refining transformers for time series forecasting, 2023.arxiv.org/abs/2206.04038
  20. [20]Peng Chen, Yingying Zhang, Yunyao Cheng, Yang Shu, Yihang Wang, Qingsong Wen, Bin Yang, and Chenjuan Guo. Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting, 2024.arxiv.org/abs/2402.05956
  21. [21]Shiyu Wang, Jiawei Li, Xiaoming Shi, Zhou Ye, Baichuan Mo, Wenze Lin, Shengtong Ju, Zhixuan Chu, and Ming Jin. Timemixer++: A general time series pattern machine for universal predictive analysis. ICLR, 2025.arxiv.org/abs/2410.16032
  22. [22]Sungwon Han, Seungeon Lee, Meeyoung Cha, Sercan O Arik, and Jinsung Yoon. Retrieval augmented time series forecasting. In Forty-second International Conference on Machine Learning, 2025.arxiv.org/abs/2505.04163
  23. [23]Hui Chen, Viet Luong, Lopamudra Mukherjee, and Vikas Singh. SimpleTM: A simple baseline for multivariate time series forecasting. In The Thirteenth International Conference on Learning Representations, 2025.
  24. [24]Cristian Challu, Kin G Olivares, Boris N Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, and Artur Dubrawski. Nhits: Neural hierarchical interpolation for time series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 6989–6997, 2023.
  25. [25]Maurice Kraus, Felix Divo, Devendra Singh Dhami, and Kristian Kersting. xLSTM-mixer: Multivariate time series forecasting by mixing via scalar memories. In Advances in Neural Information Processing Systems, volume 38, 2025.arxiv.org/abs/2410.16928
  26. [26]Keyu Tian, Yi Jiang, Zehuan Yuan, BINGYUE PENG, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.arxiv.org/abs/2404.02905
  27. [27]Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts. In The Thirteenth International Conference on Learning Representations, 2025.
  28. [28]Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Timer-XL: Long-context transformers for unified time series forecasting. In The Thirteenth International Conference on Learning Representations, 2025.arxiv.org/abs/2410.04803
  29. [29]Xiangfei Qiu, Xingjian Wu, Yan Lin, Chenjuan Guo, Jilin Hu, and Bin Yang. Duet: Dual clustering enhanced multivariate time series forecasting. In SIGKDD, pages 1185–1196, 2025.arxiv.org/abs/2412.10859
  30. [30]Siying Zhu, Jiawei Zheng, and Qianli Ma. Mr-transformer: multiresolution transformer for multivariate time series prediction. IEEE Transactions on Neural Networks and Learning Systems, 2023.DOI
  31. [31]Jiexi Liu, Meng Cao, and Songcan Chen. Musicnet: A gradual coarse-to-fine framework for irregularly sampled multivariate time series analysis. arXiv preprint arXiv:2412.01063, 2024.arxiv.org/abs/2412.01063
  32. [33]Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Mingsheng Long, and Jianmin Wang. Deep time series models: A comprehensive survey and benchmark. arXiv preprint arXiv:2407.13278, 2024.arxiv.org/abs/2407.13278
  33. [34]Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S. Jensen, Zhenli Sheng, and Bin Yang. Tfb: Towards comprehensive and fair benchmarking of time series forecasting methods. VLDB, 2024.arxiv.org/abs/2403.20150
  34. [35]Boyuan Chen, Diego Martí Monzó, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=yDo1ynArJj.arxiv.org/abs/2407.01392
  35. [36]Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In AAAI, 2021.arxiv.org/abs/2012.07436
  36. [37]Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. NeurIPS, 2021.arxiv.org/abs/2106.13008
  37. [38]Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling long-and short-term temporal patterns with deep neural networks. SIGIR, 2018.arxiv.org/abs/1703.07015
  38. [39]Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017.

Paper details

vlnc:W_pk4eha59
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents