Research map / Time series / Long-term forecasting
TimeMixer: Decomposable Multiscale Mixing 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 TimeMixer, a fully MLP-based multiscale-mixing forecaster that decomposes downsampled series into seasonal and trend parts, mixes them bottom-up and top-down respectively, and ensembles multi-scale predictors.
Problem
Long-term and short-term time series forecasting with intricate, mixed temporal variations; existing methods rely on decomposition or multiperiodicity but do not exploit multiscale information jointly in past extraction and future prediction.
Approach
Downsample the input into multiple scales, decompose each into seasonal and trend parts, mix seasonal parts bottom-up (fine-to-coarse) and trend parts top-down (coarse-to-fine) via past-decomposable-mixing blocks, then sum predictions from one linear predictor per scale (future-multipredictor-mixing).
Key results
- Long-term average MSE on Weather 0.240 vs PatchTST 0.265 (9.4% reduction) and Solar-Energy 0.216 vs 0.287 (24.7% reduction), with state-of-the-art averages across 8 long-term benchmarks under unified input length 96.
- On M4 weighted average, SMAPE 11.723, MASE 1.559, OWA 0.840, beating TimesNet (OWA 0.855).
- Efficiency: GPU memory and training speed stay low across series lengths 192-3072 compared with PatchTST, FEDformer and Autoformer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Future-Multipredictor-Mixing to combine predictions from multiple scales (ablation) | TimeMixer · M4 weighted average | OWA: 0.925 -> 0.840 | Helped |
| Add seasonal mixing (ablation) | TimeMixer · M4 weighted average | OWA: 0.962 -> 0.840 | Helped |
| Add trend mixing (ablation) | TimeMixer · M4 weighted average | OWA: 0.941 -> 0.840 | Helped |
| Use bottom-up seasonal and top-down trend mixing instead of top-down seasonal and bottom-up trend mixing (ablation) | TimeMixer · M4 weighted average | OWA: 0.954 -> 0.840 | Helped |
| Use separate seasonal and trend mixing directions rather than top-down mixing for both (ablation) | TimeMixer · M4 weighted average | OWA: 0.871 -> 0.840 | Helped |
| Use separate seasonal and trend mixing directions rather than bottom-up mixing for both (ablation) | TimeMixer · M4 weighted average | OWA: 0.859 -> 0.840 | Helped |
| Add decomposition before multiscale mixing (ablation) | TimeMixer · M4 weighted average | OWA: 0.851 -> 0.840 | Helped |
| Add decomposition before multiscale mixing (ablation) | TimeMixer · M4 weighted average | OWA: 0.850 -> 0.840 | Helped |
| Add Past-Decomposable-Mixing (ablation) | TimeMixer · M4 weighted average | OWA: 0.916 -> 0.840 | Helped |
| Increase the number of scales from 1 to 3 (ablation) | TimeMixer · ETTm1 | MSE: 0.326 -> 0.320 | Helped |
| Increase the number of scales from 3 to 4 (ablation) | TimeMixer · ETTm1 | MSE: 0.320 -> 0.321; 0.454 -> 0.454 | Mixed |
| Increase the number of PDM layers from 1 to 2 (ablation) | TimeMixer · ETTm1 | MSE: 0.328 -> 0.320 | Helped |
Surprising findings
- Under comprehensive hyperparameter search the relative gain over PatchTST is smaller than under the unified setting, as the authors acknowledge.
- Mixing trend parts bottom-up or seasonal parts top-down (opposite strategy) degrades performance, supporting the chosen direction.
Limitations
- Results under the unified setting are weaker than under per-baseline hyperparameter search; gains over PatchTST shrink.
- Number of scales M must be chosen per task (3 for long-term, 1 for short-term), requiring tuning.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Solar-Energy Electricity Traffic PEMS03 PEMS04 PEMS07 PEMS08
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Solar-Energy Electricity Traffic PEMS03 PEMS04 PEMS07 PEMS08
Metrics: MSE MAE MAPE RMSE SMAPE MASE OWA GPU memory running time
Baselines: PatchTST TimesNet Crossformer MICN FiLM DLinear FEDformer Stationary Autoformer Informer SCINet N-HiTS
Builds on: Autoformer series decomposition MLP-Mixer DLinear PatchTST
Closest papers by meaning
- TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis (ICLR 2025 · similarity 0.91)
- WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.90)
- Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting (AAAI 2025 · similarity 0.90)
- xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories (NeurIPS 2025 · similarity 0.89)
- WaveletMixer: A Multi-Resolution Wavelets Based MLP-Mixer for Multivariate Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.89)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 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.