MSMixer: Learned Multi-Scale Temporal Mixing with Complementary Linear Shortcut for Long-Term Time Series Forecasting
MSMixer: Learned Multi-Scale Temporal Mixing with Complementary Linear Shortcut for Long-Term Time Series Forecasting
Authors: Ahmed Cherif Date: 2026-05-04 Paper ID: openalex:2605.02689
Summary
MSMixer is a lightweight, channel-independent MLP architecture designed for efficient long-term time series forecasting. It utilizes three parallel branches with varied down-sampling rates and a learnable softmax gate to adaptively capture temporal dynamics across different scales. By incorporating a DLinear-based shortcut, the model effectively combines multi-scale feature learning with explicit global trend and seasonality modeling, achieving state-of-the-art results for its parameter class on ETT benchmarks.
Key Contributions
- MSMixer achieves a competitive average MSE of 0.357 on the ETT benchmarks, outperforming DLinear and NLinear.
- Introduces a multi-scale MLP architecture with parallel branches at 1x, 4x, and 16x down-sampling factors to capture multi-scale temporal patterns.
- Integrates a DLinear-based complementary shortcut to explicitly model global trend and seasonality trends alongside learned multi-scale features.
Open Questions & Future Work
Archivist Review
I have approved one open question regarding adaptive scale selection, as it addresses a fundamental design challenge in multi-scale time series modeling. The proposed open question on normalization was rejected because it suggests replacing a module with an existing method rather than exploring a deeper, unresolved theoretical or empirical bottleneck. No concepts were approved as the proposed architecture components were too tightly coupled to the specific MSMixer design or represented minor variations of existing techniques.
Approved Open Questions
- Adaptive Scale Selection Forecasters: Developing adaptive scale selection mechanisms is crucial for improving model performance on diverse datasets where fixed, heuristic-based down-sampling might not be optimal.
Rejected Candidates
- [open_question] Advanced Normalization for LTSF (
advanced-normalization-ltsf) - other: This suggests replacing a component (RevIN) with existing methods rather than identifying a deep, unresolved research bottleneck in forecasting.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.02689
- paper_id
- 2605.02689
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-forecastinglong-term-forecastinglightweight-modelsmulti-scale-modeling
- architectures
- MLP
- datasets
-
- concept_slugs
-
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-07T06:03:37Z