Research map / Time series / Long-term forecasting
TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis
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 general-purpose time series pattern machine that uses multi-resolution time imaging, dual-axis attention decomposition, and multi-scale/multi-resolution mixing to perform 8 time series tasks with state-of-the-art results.
Problem
Building a single universal time series model that captures multi-scale, multi-periodicity patterns and performs well across diverse analytical tasks (forecasting, imputation, classification, anomaly detection), where existing models are often task-specific or fail to capture overlapping temporal pa
Approach
Converts multi-scale 1D series into 2D multi-resolution time images using FFT-selected periods, decomposes each image into seasonal and trend components with column- and row-axis attention in latent space, then mixes seasonal patterns bottom-up and trend patterns top-down across scales and adaptively weights periods by FFT amplitude.
Key results
- Long-term forecasting: TimeMixer++ reports 0.165 MSE on Electricity, about 7.3% lower than iTransformer, and 0.416 MSE on Traffic.
- Univariate M4 short-term forecasting: SMAPE 11.448 and OWA 0.821, with OWA about 3.5% better than TimesNet and 11.6% better than iTransformer.
- Classification reaches 75.9% accuracy, 2.3% above TimesNet; anomaly detection reaches 87.47% average F1, 2.59% above TimesNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TimeMixer++ instead of iTransformer | Long-term forecasting · Electricity | MSE: 0.178 -> 0.165 | Helped |
| Use TimeMixer++ instead of TimeMixer | Long-term forecasting · ETT (Avg) | MSE: 0.367 -> 0.349 | Helped |
| Use TimeMixer++ instead of TimeMixer | M4 short-term forecasting · M4 (all subsets averaged) | SMAPE: 11.723 -> 11.448 | Helped |
| Use TimeMixer++ instead of TimeMixer | Multivariate short-term forecasting · PeMS (all datasets averaged) | MAE: 17.41 -> 15.91 | Helped |
| Use TimeMixer++ instead of TimesNet | Imputation · ETT | MSE: 0.079 -> 0.055 | Helped |
| Use TimeMixer++ instead of PatchTST | Few-shot forecasting with 10% training data · ETT (Avg) | MSE: 0.461 -> 0.396 | Helped |
| Use TimeMixer++ instead of PatchTST | Zero-shot forecasting · ETTh1 → ETTh2 | MSE: 0.380 -> 0.367 | Helped |
| Add channel mixing (ablation) | TimeMixer++ · Long-term forecasting (8 benchmarks; average) | MSE: 0.317 -> 0.300 | Helped |
| Add time image decomposition (ablation) | TimeMixer++ · Long-term forecasting (8 benchmarks; average) | MSE: 0.329 -> 0.300 | Helped |
| Add multi-scale mixing (ablation) | TimeMixer++ · Long-term forecasting (8 benchmarks; average) | MSE: 0.320 -> 0.300 | Helped |
| Add multi-resolution mixing (ablation) | TimeMixer++ · Long-term forecasting (8 benchmarks; average) | MSE: 0.316 -> 0.300 | Helped |
| Add channel mixing (ablation) | TimeMixer++ · Multivariate short-term forecasting (PEMS datasets) | MAPE: 12.17 -> 10.35 | Helped |
Surprising findings
- Channel-independence is abandoned in favor of channel mixing via variate-wise self-attention at the coarsest scale, which the authors find improves performance on high-dimensional datasets such as PEMS.
Limitations
- Limitations are deferred to an appendix section not included in the provided text.
- Hyperparameter sensitivity (number of scales, layers, input length, top-K periods) is required to be tuned per task, and the scale count M is chosen according to series length.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic Solar-Energy Exchange M4 PEMS03 PEMS04
Benchmarks: ETT Weather Electricity Traffic Solar-Energy Exchange M4 PEMS03 PEMS04 PEMS07 PEMS08 ETT imputation
Metrics: MSE MAE SMAPE MASE OWA MAPE RMSE accuracy F1
Baselines: TimeMixer iTransformer PatchTST Crossformer TiDE TimesNet DLinear SCINet FEDformer Stationary Autoformer MICN
Builds on: TimesNet TimeMixer PatchTST iTransformer
Closest papers by meaning
- TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (ICLR 2024 · similarity 0.91)
- Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.88)
- WaveletMixer: A Multi-Resolution Wavelets Based MLP-Mixer for Multivariate Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.87)
- TSLANet: Rethinking Transformers for Time Series Representation Learning (ICML 2024 · similarity 0.87)
- Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting (AAAI 2025 · similarity 0.87)
- Time-MMD: Multi-Domain Multimodal Dataset for Time Series Analysis (NeurIPS 2024 · 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.