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

Fanda Fan, Kuiye Ding, Liming Mao, Bingsong Wang, Yao Wang, Xiaorui Wang, Ruijie Jian, Zhipeng Liu, Luqi Gong, Zhenghua Lu, Chunjie Luo, Jianfeng ZhanarXiv 2026 · arXiv

TL;DR

MSAR (Modular Scale-wise Autoregressive Framework) is a model-agnostic "wrapper" that improves time series forecasting by predicting progressively from coarse to fine temporal resolutions. By aligning inputs and outputs at each scale and using global trends to guide local predictions, it achieves superior accuracy and up to 2x faster inference across diverse architectures like Transformers and MLPs.

What Problem Does This Paper Address?

Time series data is inherently multi-scale: a retail dataset contains long-term seasonal trends (yearly/monthly) and short-term fluctuations (daily/hourly). Most current deep learning models suffer from two main issues:

  1. Single-scale Modeling: They try to capture all these heterogeneous patterns in a single representation, leading to "scale interference" where noise at one level obscures signals at another.
  2. Scale Misalignment: Even models that use multi-scale inputs usually produce a single-scale output. This creates a mismatch that complicates the learning process and limits the model's ability to leverage hierarchical information.

The authors argue that for a model to truly understand a time series, it should predict the "big picture" first and then progressively fill in the details.

Methods

The core of the paper is the Modular Scale-wise Autoregressive Framework (MSAR). Unlike traditional autoregressive models that predict the next time step, MSAR predicts the next scale.

Key components include:

  • Scale-wise Aligned Modeling: The framework decomposes the input into multiple resolutions (e.g., 1x, 2x, 4x downsampled). Crucially, it forces the model to predict the output at the same resolution as the input for that specific scale.
  • Scale-wise Autoregression: Forecasting happens hierarchically. The coarsest scale is predicted first. This prediction is then upsampled and passed as "contextual guidance" to the next finer scale.
  • Imputation-based Refinement: Instead of just adding components together, finer-scale modules act as "imputation" engines. They take the coarse prediction and the current-scale input to refine and "fill in" the higher-frequency details.
  • Modular Design: MSAR is model-agnostic. You can "plug in" almost any backbone—be it a Linear model (DLinear), a Transformer (PatchTST, iTransformer), or a CNN (TimesNet).
  • Two-Stage Training: To combat "exposure bias" (where errors in coarse predictions snowball into fine-scale disasters), the authors use a teacher-forcing stage followed by a joint training stage.

Datasets

The authors evaluated MSAR on a massive suite of 14 real-world multivariate time series datasets, covering:

  • Long-term Forecasting: ETT (4 subsets), ECL (Electricity), Exchange, Traffic, Weather, Solar-Energy, and Wind.
  • Short-term Forecasting: PEMS (4 subsets of traffic data). These represent standard benchmarks in the field, allowing for direct comparison with state-of-the-art (SOTA) models.

Key Insights and Ideas

  • Alignment is more important than "more data": One of the most striking findings (Table 1) is that a model trained on downsampled data to predict downsampled targets often outperforms a model given full-resolution data to predict the same downsampled targets. Alignment reduces learning complexity.
  • The "Skeleton" Effect: Coarse-scale predictions provide a stable structural "skeleton." This prevents the model from getting lost in high-frequency noise, which is particularly beneficial for long-horizon forecasting.
  • Efficiency via Decoupling: By breaking the problem into scales, the model can often process shorter effective sequence lengths at the finer scales, leading to significant inference speedups.
  • Synergy of Components: Ablation studies showed that neither multi-scale inputs nor scale-wise autoregression is enough on its own; the "magic" happens when they are combined with strict input-output alignment.

Key Results

  • Consistent Accuracy Gains: MSAR improved performance across nearly all model-dataset pairs. For example, DLinear’s MSE on the Exchange dataset dropped from 0.524 to 0.372—a massive improvement for a simple linear backbone.
  • Inference Speedup: For models like PatchTST and DLinear, MSAR reduced inference latency by nearly a factor of two. This is because the scale-wise approach allows for more efficient parallel decoding compared to traditional step-by-step autoregression.
  • Robustness to Context Length: MSAR effectively utilizes longer "look-back" windows. While many models plateau or overfit as you give them more history, MSAR's decoupled structure allows it to extract relevant long-term trends more effectively.

Limitations and Future Work

  • Error Propagation: As with any autoregressive process, if the coarsest scale makes a major mistake, it can be difficult for finer scales to fully correct it.
  • Training Complexity: The two-stage training process (Teacher Forcing + Joint Training) increases the total training time compared to single-stage baselines.
  • Future Directions: The authors suggest exploring "residual corrections" to better handle inherited errors and investigating how MSAR scales with massive "foundation model" pretraining. They also note that the choice of interval sizes (how much you downsample) currently requires some manual tuning.

Summary updated

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.
BibTeX
@misc{fan2026msar,
  author = {Fanda Fan and Kuiye Ding and Liming Mao and Bingsong Wang and Yao Wang and Xiaorui Wang and Ruijie Jian and Zhipeng Liu and Luqi Gong and Zhenghua Lu and Chunjie Luo and Jianfeng Zhan},
  title = {{MSAR: Next-Scale Autoregressive Forecasting for Time Series via Modular Multi-Scale Decoupling}},
  year = {2026},
  eprint = {2611.20714},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20714},
}
Figure