Research map / Time series / Long-term forecasting
Parsimony or Capability? Decomposition Delivers Both in Long-term 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 SSCNN, a decomposition-based long-term forecasting network with attention-based selection maps for long-term, seasonal, short-term and spatial components, using over 99% fewer parameters than SOTA while matching or beating accuracy.
Problem
Long-term time series forecasting models have grown to millions or billions of parameters without proportionate accuracy gains; the paper asks whether a parsimonious model can match SOTA capability.
Approach
Decompose the series into long-term, seasonal, short-term and spatial structured components, each extracted and extrapolated by a learned selection/attention map (T-AttnNorm, S-AttnNorm), with a polynomial regression layer fusing components; argues analytically that decomposition separates components more cleanly than patching while needing far fewer parameters.
Key results
- On Electricity, SSCNN achieves MSE 0.118 vs iTransformer 0.121 and PatchTST 0.125 (averaged over horizons), and on Traffic MSE 0.338 vs 0.343 and 0.349.
- SSCNN uses about 99% fewer parameters than most competing methods and 87% fewer than DLinear at forward window 336 on ECL.
- Ablation shows removing the seasonal component hurts most; adding an FCN ('w FCN') gives no benefit.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the selection mechanism to SCNN (ablation) | SCNN · ECL and Traffic | MSE and MAE: approximate 8% improvement | Helped |
| Add the selection mechanism (ablation) | SCNN · ETTh1, ETTh2, ETTm1, ETTm2, and Weather | MSE and MAE | No clear effect |
| Remove the long-term component (ablation) | SSCNN · ECL and Traffic | forecasting accuracy | Hurt |
| Remove the seasonal component (ablation) | SSCNN · ECL and Traffic | forecasting accuracy | Hurt |
| Remove the short-term component (ablation) | SSCNN · ECL and Traffic | forecasting accuracy | Hurt |
| Remove the spatial component (ablation) | SSCNN · ECL and Traffic | forecasting accuracy | Hurt |
| Remove temporal attention maps (ablation) | SSCNN · ECL and Traffic | forecasting accuracy | Hurt |
| Remove spatial attention maps (ablation) | SSCNN · ECL and Traffic | forecasting accuracy | Hurt |
| Insert an additional fully-connected network between backward and forward variables (ablation) | SSCNN · ECL and Traffic | forecasting accuracy | No clear effect |
| Use SSCNN instead of iTransformer | Long-term forecasting benchmarks · Seven evaluated datasets | MSE | Helped |
Surprising findings
- SSCNN is slower in FLOPs than iTransformer because it processes the full sequence at each layer rather than compressing with fully connected layers.
- Inter-channel conditional correlation via the spatial component improves results, contrary to prior evidence that channel-dependence harms performance.
Limitations
- Computational cost (FLOPs) is higher than iTransformer and scales with the sum of input and output lengths.
- Spatial component is disabled on datasets without spatial correlations, so the model is not uniformly applied.
- Hyperparameters such as cycle length must be configured to match the data's period, with notable performance drops when misaligned.
Experimental setup
Datasets: Weather Traffic Electricity ETTh1 ETTh2 ETTm1 ETTm2
Benchmarks: Weather Traffic Electricity ETTh1 ETTh2 ETTm1 ETTm2
Metrics: MSE MAE
Baselines: SCNN iTransformer TimeMixer PatchTST TimesNet Crossformer DLinear Autoformer
Builds on: SCNN DLinear PatchTST iTransformer
Closest papers by meaning
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.89)
- SparseTSF: Modeling Long-term Time Series Forecasting with *1k* Parameters (ICML 2024 · similarity 0.88)
- Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.87)
- A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle (ICML 2025 · similarity 0.87)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.87)
- Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting (AAAI 2025 · similarity 0.87)
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.