Research map / Time series / Long-term forecasting
Amplifier: Bringing Attention to Neglected Low-Energy Components in 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 an energy amplification technique (spectrum-flipping amplify/restore blocks) plus the Amplifier model (seasonal-trend forecaster with a semi-channel interaction block) to stop forecasters from neglecting low-energy frequency components.
Problem
Deep time series forecasting models focus on high-energy frequency components and neglect low-energy components, which the paper argues are indispensable for accurate prediction.
Approach
Amplify low-energy frequency components by adding a flipped copy of the spectrum (X + X'), so the model sees balanced energy, then remove the added spectrum from the output via an energy restoration block; combine this with a seasonal-trend forecaster and a commonality/specificity channel-interaction block.
Key results
- Amplifier with lookback 96 achieves best MSE/MAE on 6 of 8 datasets (ETTm1 MSE 0.381, ETTm2 0.276, ECL 0.171) and is weaker on Traffic (MSE 0.482 vs. Fredformer 0.431).
- Adding the energy amplification technique improves Amplifier by 5.643% MSE on ETTh1 and 15.100% MSE on Weather (ablation).
- Plugging energy amplification into iTransformer, SparseTSF and DLinear yields average gains of about 9.8% MSE for Transformers and 3.6% MSE for linear models on ETTm1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add energy amplification technique (ablation) | Amplifier · ETTh1 | MSE: 0.393 -> 0.371; 5.643% | Helped |
| add energy amplification technique (ablation) | Amplifier · ETTh1 | MAE: 0.407 -> 0.392; 3.613% | Helped |
| add energy amplification technique (ablation) | Amplifier · ETTm2 | MSE: 0.182 -> 0.176; 3.484% | Helped |
| add energy amplification technique (ablation) | Amplifier · ETTm2 | MAE: 0.266 -> 0.258; 2.973% | Helped |
| add energy amplification technique (ablation) | Amplifier · Weather | MSE: 0.184 -> 0.156; 15.100% | Helped |
| add energy amplification technique (ablation) | Amplifier · Weather | MAE: 0.230 -> 0.205; 10.923% | Helped |
| add SCI block (ablation) | Amplifier · Electricity | MSE: 5.769%, 5.988%, 7.839%, 10.318% | Helped |
| add SCI block (ablation) | Amplifier · Electricity | MAE: 1.619%, 1.953%, 3.237%, 5.219% | Helped |
| add SCI block (ablation) | Amplifier · Traffic | MSE: 9.000%, 1.426%, 5.838%, 3.807% | Helped |
| add SCI block (ablation) | Amplifier · Traffic | MAE: 9.063%, 0.230%, 3.067%, 4.817% | Helped |
| add energy amplification technique | Transformer-based foundational models · ETTm1 | MSE: 9.837% | Helped |
| add energy amplification technique | Transformer-based foundational models · ETTm1 | MAE: 4.236% | Helped |
Surprising findings
- Amplifier performs poorly on Traffic, which the authors attribute to strong periodicity being carried by high-energy components that the technique does not target.
Limitations
- Not the lightest model; larger parameters and slower training than linear baselines.
- Weaker on Traffic, a strongly periodic dataset where high-energy components dominate.
Experimental setup
Datasets: ETTm1 ETTm2 ETTh1 ETTh2 Electricity Exchange Traffic Weather
Benchmarks: ETTm1 ETTm2 ETTh1 ETTh2 Electricity Exchange Traffic Weather
Metrics: MSE MAE
Baselines: RLinear DLinear SparseTSF FreTS FITS Fredformer iTransformer PatchTST Stationary TimesNet
Builds on: DLinear FreTS FITS iTransformer PatchTST
Closest papers by meaning
- FilterNet: Harnessing Frequency Filters for Time Series Forecasting (NeurIPS 2024 · similarity 0.85)
- SEMPO: Lightweight Foundation Models for Time Series Forecasting (NeurIPS 2025 · similarity 0.84)
- A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle (ICML 2025 · similarity 0.84)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.84)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.84)
- SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention (ICML 2024 · similarity 0.84)
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.