TimeMar: Multi-Scale Autoregressive Modeling for Unconditional Time Series Generation

TL;DR

TimeMAR tackles unconditional time-series generation by explicitly modeling multi-scale temporal structure and structural heterogeneity. It combines a trend–seasonal decomposition via MoE, a dual-path multi-scale VQ-VAE encoder, and a coarse-guided reconstruction strategy with a GPT-style autoregressive model operating on discrete tokens. Across six datasets, TimeMAR achieves superior generation fidelity and long-horizon ability with markedly fewer parameters than competing methods, and ablations confirm the value of each component. The approach offers a scalable, privacy-preserving path for realistic synthetic time-series data in Web/Mobile/ WoT contexts and lays groundwork for conditional extensions and missing-data handling.

Abstract

Generative modeling offers a promising solution to data scarcity and privacy challenges in time series analysis. However, the structural complexity of time series, characterized by multi-scale temporal patterns and heterogeneous components, remains insufficiently addressed. In this work, we propose a structure-disentangled multiscale generation framework for time series. Our approach encodes sequences into discrete tokens at multiple temporal resolutions and performs autoregressive generation in a coarse-to-fine manner, thereby preserving hierarchical dependencies. To tackle structural heterogeneity, we introduce a dual-path VQ-VAE that disentangles trend and seasonal components, enabling the learning of semantically consistent latent representations. Additionally, we present a guidance-based reconstruction strategy, where coarse seasonal signals are utilized as priors to guide the reconstruction of fine-grained seasonal patterns. Experiments on six datasets show that our approach produces higher-quality time series than existing methods. Notably, our model achieves strong performance with a significantly reduced parameter count and exhibits superior capability in generating high-quality long-term sequences. Our implementation is available at https://anonymous.4open.science/r/TimeMAR-BC5B.

Figures (7)

• Figure 1: Multi-Scale Representation of Time Series and Structural Decomposition. Left: Multi-Scale Representation. The series is further factorized across multiple resolutions, allowing hierarchical modeling. Right: Seasonal-Trend Decomposition. The input time series (gray) is decomposed into three interpretable components: underlying trend (top), fine-grained seasonality (middle), and coarse-grained seasonality (bottom).
• Figure 2: Scaling behavior of different model families on the Context-FID metric on Energy.
• Figure 3: Overview of the proposed TimeMAR framework. (a) Architecture of the Multi-Scale VQ-VAE Module, which decomposes the input sequence and encodes components into multi-scale discrete representations using coarse and fine encoders. (b) Sequence Modeling with Transformer, where a GPT-style autoregressive model learns token dependencies. (c) Time-Series Reconstruction from Tokens, where generated tokens are decoded into continuous sequences.
• Figure 4: Visualization of synthesized time series by t-SNE (1st row), PCA (2nd row), and Kernel density estimation (3rd row)
• Figure 5: Visualization of the Interpretable Generation Process. Left: Energy Dataset; Right: ETTh Dataset.