Research map / Time series / Long-term forecasting
FilterNet: Harnessing Frequency Filters for 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 FilterNet, a lightweight forecaster built from two learnable frequency filters (plain and contextual shaping) wrapped in instance normalization and an FFN, evaluated on eight long-term forecasting benchmarks.
Problem
Transformer-based time series forecasters are vulnerable to high-frequency signals, costly in computation, and underuse the full frequency spectrum when predicting long series.
Approach
Replace linear and attention mappings with learnable frequency filters applied in the Fourier domain: a plain filter with a randomly initialized learnable kernel (shared or per-channel), and a contextual filter whose response is generated from the input spectrum, so the model acts as an all-pass filter that can surrogate MLP and attention behaviour.
Key results
- PaiFilter achieves best MSE on ETTm1 (0.318 at horizon 96) and ETTh1 horizon 720 sits near top across baselines under lookback 96.
- TexFilter attains the best ECL MSE at horizons 96 through 720 (e.g., 0.147 at horizon 96), with iTransformer best on Traffic.
- FilterNet's efficiency is reported as similar to DLinear and faster than Transformer models on Exchange and Electricity, with complexity O(log L) claimed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FilterNet (PaiFilter) instead of the strongest listed baseline | ETTm1 · ETTm1, prediction length 96 | MSE: 0.316 -> 0.318 | Mixed |
| Use FilterNet (PaiFilter) instead of PatchTST | ETTm2 · ETTm2, prediction length 96 | MSE: 0.175 -> 0.174 | Helped |
| Use FilterNet (PaiFilter) instead of FEDformer | ETTh1 · ETTh1, prediction length 96 | MSE: 0.376 -> 0.375 | Helped |
| Use FilterNet (PaiFilter) instead of iTransformer | ETTh2 · ETTh2, prediction length 96 | MSE: 0.297 -> 0.292 | Helped |
| Use FilterNet (TexFilter) instead of iTransformer | ECL · ECL, prediction length 96 | MSE: 0.148 -> 0.147 | Helped |
| Use FilterNet (PaiFilter) instead of FITS | Exchange · Exchange, prediction length 96 | MSE: 0.084 -> 0.083 | Helped |
| Use FilterNet (TexFilter) instead of iTransformer | Traffic · Traffic, prediction length 96 | MSE: 0.395 -> 0.430 | Hurt |
| Use FilterNet (TexFilter) instead of iTransformer | Weather · Weather, prediction length 96 | MSE: 0.174 -> 0.162 | Helped |
| Add instance normalization and inverse normalization (ablation) | FilterNet · ETTh1, ETTm1, and Electricity, prediction length 96 | MSE | Helped |
| Add the frequency filter block (ablation) | FilterNet · ETTh1, ETTm1, and Electricity, prediction length 96 | MSE | Helped |
| Add the feed-forward network (ablation) | FilterNet · ETTh1, ETTm1, and Electricity, prediction length 96 | MSE | Helped |
| Use shared filters instead of channel-unique filters (ablation) | FilterNet · ETTh1 and Exchange, prediction lengths 96, 192, 336, and 720 | MSE: 0.382 -> 0.375; 0.091 -> 0.083 | Helped |
Surprising findings
- iTransformer, despite strong overall results, performs best on Traffic, which the authors attribute to Traffic energy concentrated in low frequencies.
- Shared (universal) plain filters outperform per-channel (individual) filters on ETTh1 and Exchange.
Limitations
- The paper reports no statistical test details beyond a general 95% significance claim.
- Results rely on RevIN-style instance normalization and tuned FFN hidden size, so gains depend on these components.
- Plain filter underperforms on larger datasets such as Electricity and Traffic.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Weather Traffic Exchange
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Weather Traffic Exchange
Metrics: MSE MAE
Baselines: iTransformer PatchTST FEDformer TimesNet DLinear RLinear FITS FreTS
Builds on: RevIN DLinear PatchTST FreTS
Closest papers by meaning
- FilterTS: Comprehensive Frequency Filtering for Multivariate Time Series Forecasting (AAAI 2025 · similarity 0.91)
- TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting (ICML 2025 · similarity 0.88)
- TimeKAN: KAN-based Frequency Decomposition Learning Architecture for Long-term Time Series Forecasting (ICLR 2025 · similarity 0.88)
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.88)
- Unified Training of Universal Time Series Forecasting Transformers (ICML 2024 · similarity 0.88)
- TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.88)
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.