Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DMamba: Decomposition-enhanced Mamba for Time Series Forecasting

Ruxuan Chen, Fang Sun

TL;DR

DMamba introduces a decomposition-aware, dual-stream forecasting architecture for non-stationary multivariate time series. By applying EMA-based decomposition, it separates a low-dimensional trend from high-frequency seasonal residuals and processes them with a lightweight trend MLP and a Mamba-based seasonal backbone, respectively. The approach is backed by RevIN normalization and an Arctan Weighted Loss to stabilize training and prioritize near-term accuracy. Empirical results on diverse benchmarks (ETT, Weather, PEMS) show state-of-the-art performance and favorable efficiency, with ablations validating the architectural design and decomposition strategy.

Abstract

State Space Models (SSMs), particularly Mamba, have shown potential in long-term time series forecasting. However, existing Mamba-based architectures often struggle with datasets characterized by non-stationary patterns. A key observation from time series theory is that the statistical nature of inter-variable relationships differs fundamentally between the trend and seasonal components of a decomposed series. Trend relationships are often driven by a few common stochastic factors or long-run equilibria, suggesting that they reside on a lower-dimensional manifold. In contrast, seasonal relationships involve dynamic, high-dimensional interactions like phase shifts and amplitude co-movements, requiring more expressive modeling. In this paper, we propose DMamba, a novel forecasting model that explicitly aligns architectural complexity with this component-specific characteristic. DMamba employs seasonal-trend decomposition and processes the components with specialized, differentially complex modules: a variable-direction Mamba encoder captures the rich, cross-variable dynamics within the seasonal component, while a simple Multi-Layer Perceptron (MLP) suffices to learn from the lower-dimensional inter-variable relationships in the trend component. Extensive experiments on diverse datasets demonstrate that DMamba sets a new state-of-the-art (SOTA), consistently outperforming both recent Mamba-based architectures and leading decomposition-based models.

DMamba: Decomposition-enhanced Mamba for Time Series Forecasting

TL;DR

DMamba introduces a decomposition-aware, dual-stream forecasting architecture for non-stationary multivariate time series. By applying EMA-based decomposition, it separates a low-dimensional trend from high-frequency seasonal residuals and processes them with a lightweight trend MLP and a Mamba-based seasonal backbone, respectively. The approach is backed by RevIN normalization and an Arctan Weighted Loss to stabilize training and prioritize near-term accuracy. Empirical results on diverse benchmarks (ETT, Weather, PEMS) show state-of-the-art performance and favorable efficiency, with ablations validating the architectural design and decomposition strategy.

Abstract

State Space Models (SSMs), particularly Mamba, have shown potential in long-term time series forecasting. However, existing Mamba-based architectures often struggle with datasets characterized by non-stationary patterns. A key observation from time series theory is that the statistical nature of inter-variable relationships differs fundamentally between the trend and seasonal components of a decomposed series. Trend relationships are often driven by a few common stochastic factors or long-run equilibria, suggesting that they reside on a lower-dimensional manifold. In contrast, seasonal relationships involve dynamic, high-dimensional interactions like phase shifts and amplitude co-movements, requiring more expressive modeling. In this paper, we propose DMamba, a novel forecasting model that explicitly aligns architectural complexity with this component-specific characteristic. DMamba employs seasonal-trend decomposition and processes the components with specialized, differentially complex modules: a variable-direction Mamba encoder captures the rich, cross-variable dynamics within the seasonal component, while a simple Multi-Layer Perceptron (MLP) suffices to learn from the lower-dimensional inter-variable relationships in the trend component. Extensive experiments on diverse datasets demonstrate that DMamba sets a new state-of-the-art (SOTA), consistently outperforming both recent Mamba-based architectures and leading decomposition-based models.
Paper Structure (25 sections, 18 equations, 3 figures, 5 tables)

This paper contains 25 sections, 18 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Transformer-based LTSF Models
  4. Mamba for Time Series Forecasting
  5. Decomposition and Exponential Smoothing
  6. Preliminaries
  7. Problem Definition
  8. State Space Models
  9. Methodology
  10. Reversible Instance Normalization
  11. EMA-based Decomposition
  12. Dual-stream Network
  13. Seasonal Stream (Mamba Backbone)
  14. Trend Stream
  15. Aggregation and Denormalization
  16. ...and 10 more sections

Figures (3)

  • Figure 1: The Architecture of DMamba.
  • Figure 2: The Architecture of the Mamba Block.
  • Figure 3: Efficiency trade-off on ETTh1 and PEMS03 with a look-back window of $L=96$. DMamba (bottom-left) achieves the best balance of low error (Avg MSE), fast inference, and low computational cost (FLOPs, bubble size).