Research map / Time series / Long-term forecasting
FilterTS: Comprehensive Frequency Filtering for Multivariate Time Series Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FilterTS, a frequency-domain multivariate forecaster with a dynamic cross-variable filtering module and a static global band-pass filtering module, evaluated on eight long-term forecasting benchmarks.
Problem
Multivariate long-term time series forecasting struggles to capture intricate periodic and trend components, including stable frequency patterns and variable frequency components shared across variables.
Approach
Treats each variable's frequency spectrum as a dynamic FIR filter applied to other variables within each lookback window (cross-variable filtering), combined with static band-pass filters built from top-K dominant frequencies computed over the full training set, all operating in the frequency domain where convolution becomes multiplication.
Key results
- Averaged over horizons, FilterTS achieves best MSE on 6 of 8 datasets, e.g., ETTh1 avg MSE 0.433 vs 0.453 for TimeMixer (lookback 96).
- Ablation on ETTm1: removing the static global filtering module raises avg MSE from 0.385 to 0.392; removing both modules raises it to 0.405.
- On Traffic, FilterTS uses 3.81GB memory versus 12.09GB for iTransformer, 18.98GB for PatchTST and 30.76GB for TimeMixer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Static Global Filtering (ablation) | FilterTS · ETTm1, Weather, Electricity | MSE: 0.392 -> 0.385; 0.250 -> 0.244; 0.197 -> 0.180 | Helped |
| add Dynamic Cross-Variable Filtering (ablation) | FilterTS · ETTm1, Weather, Electricity | MSE: 0.390 -> 0.385; 0.250 -> 0.244; 0.184 -> 0.180 | Helped |
| add both filtering modules (ablation) | FilterTS · ETTm1, Weather, Electricity | MSE: 0.405 -> 0.385; 0.267 -> 0.244; 0.208 -> 0.180 | Helped |
| use Dynamic Cross-Variable Filtering instead of MLP (ablation) | FilterTS · ETTm1, Weather, Electricity | MSE: 0.387 -> 0.385; 0.248 -> 0.244; 0.204 -> 0.180 | Helped |
| use Dynamic Cross-Variable Filtering instead of attention (ablation) | FilterTS · ETTm1, Weather, Electricity | MSE: 0.392 -> 0.385; 0.254 -> 0.244; 0.181 -> 0.180 | Helped |
Surprising findings
- FilterTS is weaker than iTransformer on Traffic and Electricity, which the authors attribute to the high variable count and a static weight matrix.
Limitations
- Weaker than iTransformer on high-dimensional datasets (Traffic, Electricity), attributed to the static inter-variable weight matrix.
- Static filters are fixed after training-set analysis and may not adapt to distribution shift.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Exchange Traffic Weather
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Exchange Traffic Weather
Metrics: MSE MAE training time memory usage
Baselines: TimeMixer iTransformer PatchTST Crossformer TimesNet FreTS
Builds on: FreTS iTransformer FITS instance normalization
Closest papers by meaning
- FilterNet: Harnessing Frequency Filters for Time Series Forecasting (NeurIPS 2024 · similarity 0.91)
- HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.88)
- MSGNet: Learning Multi-Scale Inter-series Correlations for Multivariate Time Series Forecasting (AAAI 2024 · similarity 0.88)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.87)
- Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling (ICML 2024 · similarity 0.87)
- Revisiting Attention for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.