MSMixer: Learned Multi-Scale Temporal Mixing with Complementary Linear Shortcut for Long-Term Time Series Forecasting

Problem
This paper addresses the limitations of existing lightweight MLP-based models in long-term time series forecasting, which typically operate at a single temporal resolution. This restricts their ability to capture diverse temporal patterns, including rapid oscillations, medium-range periodicities, and slowly evolving macro-trends. The authors propose MSMixer, a novel architecture designed to effectively model these multi-scale patterns. This work is presented as a preprint and has not yet undergone peer review.

Method
MSMixer introduces a channel-independent multi-scale MLP architecture characterized by three key innovations: (i) three parallel scale branches with down-sample factors of 1x, 4x, and 16x, each equipped with independent MLP blocks, (ii) a learnable softmax gate that dynamically weights the outputs from these branches, and (iii) a DLinear complementary shortcut that integrates full-window trend and seasonality context. The model is lightweight, containing only 112K parameters at a horizon of H=96, and operates with O(T) complexity, making it computationally efficient for long-term forecasting tasks.

Results
MSMixer was evaluated on four ETT benchmarks using standard chronological splits and three random seeds. It achieved the lowest average Mean Squared Error (MSE) of 0.357 among lightweight models, outperforming DLinear (0.386, a 7.4% improvement) and NLinear (0.365, a 2.1% improvement), winning 12 out of 16 configurations. When compared to five Transformer-based baselines, MSMixer achieved the best or second-best MSE in 9 out of 16 configurations while utilizing 5x fewer parameters than PatchTST. These results demonstrate the effectiveness of the proposed architecture in capturing multi-scale temporal dynamics.

Limitations
The authors acknowledge that while MSMixer shows promising results, its performance may vary across different datasets and forecasting horizons, which could limit generalizability. Additionally, the reliance on a fixed look-back window may not capture all relevant temporal dependencies in certain applications. The paper does not address potential overfitting issues or the model’s performance in real-world scenarios, which could be critical for practical deployment.

Why it matters
The development of MSMixer has significant implications for the field of time series forecasting, particularly in applications requiring the integration of multiple temporal scales. By demonstrating that a lightweight MLP architecture can outperform more complex models, this work encourages further exploration of efficient architectures for time series analysis. The findings may inspire future research to investigate additional enhancements in model design, such as incorporating more sophisticated gating mechanisms or exploring alternative shortcut connections. Overall, MSMixer represents a step forward in making long-term forecasting more accessible and effective.

Authors: Ahmed Cherif
Source: arXiv:2605.02689
URL: https://arxiv.org/abs/2605.02689v1