FilterTS: Comprehensive Frequency Filtering for Multivariate Time Series Forecasting

Yulong Wang, Yushuo Liu, Xiaoyi Duan, Kai WangAAAI 2025 · DOI

TL;DR

This paper introduces FilterTS, a novel approach to multivariate time series forecasting that leverages frequency domain analysis to extract and emphasize crucial periodic components. By employing specialized static and dynamic filtering modules, FilterTS significantly improves prediction accuracy and computational efficiency compared to existing methods, particularly for datasets with a high number of variables.

What Problem Does This Paper Address?

Multivariate time series forecasting is critical across many domains, but existing models often struggle to accurately capture the complex periodic and trend components within the data. While some methods have begun to explore frequency domain analysis, they often fail to differentiate the importance of various frequency components, leading to potential overfitting or neglecting crucial a priori periodic information and inter-variable dependencies. This paper addresses the challenge of effectively extracting and leveraging these frequency components, especially those that dynamically change or are shared across variables, to enhance forecasting performance.

Methods

FilterTS is built around a novel architecture that operates primarily in the frequency domain. The core innovation lies in its two filtering modules:

  • Dynamic Cross-Variable Filtering Module (DCFilter): This module treats each time series variable as a potential filter for others. It dynamically identifies and extracts shared frequency components within a given lookback window, adapting to real-time fluctuations. This is achieved by evaluating frequency component amplitudes and using a threshold to focus on significant ones, effectively capturing variable frequency components.
  • Static Global Filtering Module (SGFilter): This module identifies stable, dominant frequency components present throughout the entire training dataset. It constructs band-pass filters targeting these high-amplitude frequencies, ensuring the capture of consistent periodic patterns.

The model converts time-domain data to the frequency domain using the Fast Fourier Transform (FFT), which allows for efficient convolution operations to be performed as point-wise multiplications. The outputs from the DCFilter and SGFilter are then aggregated and transformed back into the time domain to generate forecasts. This frequency-domain approach significantly improves computational efficiency.

Datasets

The paper evaluates FilterTS on eight widely-recognized multivariate time series forecasting datasets:

  • ETT Datasets: ETTh1, ETTh2, ETTm1, ETTm2 (energy domain, hourly and 15-minute intervals).
  • Electricity: Electricity consumption metrics (energy domain, 15-minute intervals).
  • Exchange Rate: Daily exchange rates for major currencies (economic domain).
  • Traffic: Traffic flow and occupancy rates (transportation domain, hourly intervals).
  • Weather: Meteorological data (climatology domain, 10-minute intervals).

These datasets span diverse domains and variable counts, ensuring a comprehensive evaluation.

Key Insights and Ideas

  • Frequency Domain is Key for Complex Patterns: The paper strongly advocates for the frequency domain as a superior space for identifying and isolating complex periodic and trend components in multivariate time series.
  • Dynamic Filtering for Adaptability: The Dynamic Cross-Variable Filtering Module is a crucial innovation, allowing the model to adapt to changing frequency patterns and capture inter-variable dependencies that evolve over time. This moves beyond static frequency analysis.
  • Shared Frequencies are Predictive: The success of the DCFilter highlights that synchronized frequency components across different variables are highly predictive of future behavior.
  • Static Components Provide a Stable Baseline: The Static Global Filtering Module ensures that the model doesn't miss fundamental, stable periodicities that are present consistently in the data.
  • Computational Efficiency through Frequency Domain: By converting convolutions to multiplications in the frequency domain, FilterTS achieves significant speedups, making it practical for large-scale datasets.

Key Results

FilterTS demonstrates superior performance across the evaluated datasets:

  • Accuracy: FilterTS achieved the best performance on 6 out of 8 datasets and the second-best on the remaining 2, averaging a reduction in MSE by 4.24% compared to PatchTST and 3.69% compared to TimeMixer.
  • Efficiency: The model significantly outperforms others in terms of memory usage and training speed. For instance, on the Traffic dataset, FilterTS required only 3.81GB of memory compared to iTransformer's 12.09GB and TimeMixer's 30.76GB.
  • Ablation Studies: Removing either the Static Global Filtering Module or the Dynamic Cross-Variable Filtering Module led to performance degradation, underscoring the importance of both components. Replacing the dynamic module with a simple MLP or attention mechanism also resulted in inferior performance, suggesting the filtering approach is more effective.

Limitations and Future Work

  • High-Dimensional Datasets: The paper notes that FilterTS's performance on datasets with a very high number of variables (like Traffic) can be slightly weaker due to challenges in capturing complex inter-variable relationships with the static weight matrix. Models with sophisticated attention mechanisms might handle these better.
  • Hyperparameter Tuning: The paper includes a sensitivity analysis, highlighting the importance of carefully tuning hyperparameters like the quantile threshold for dynamic filters, the number of static filters, and their bandwidth for optimal performance.
  • Future Work: While not explicitly detailed, the success of FilterTS suggests future work could explore more advanced frequency decomposition techniques, adaptive bandwidth selection, and further integration of complex number theory for even more nuanced frequency analysis. The paper also implies that refining frequency components is a promising avenue for advancing time series forecasting.

Summary updated

BibTeX
@inproceedings{wang2025filterts,
  author = {Yulong Wang and Yushuo Liu and Xiaoyi Duan and Kai Wang},
  title = {{FilterTS: Comprehensive Frequency Filtering for Multivariate Time Series Forecasting}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i20.35438},
}
Figure