TimesNet
Folds 1D series into 2D images by period, then lets vision-style kernels do the temporal modeling.
About
TimesNet (Wu, Hu, Liu, Zhou, Wang, Long, ICLR 2023) starts from the observation that real series are multi-periodic, and models that in the most literal way possible: FFT-detected periods fold the 1D series into 2D tensors, where columns hold intra-period variation and rows hold inter-period variation, and a parameter-efficient inception block with 2D convolutions models both at once. One backbone then serves five tasks: long-term and short-term forecasting, imputation, classification, and anomaly detection.
The authors report consistent state-of-the-art across all five tasks against 15-plus baselines in their benchmark. The original thuml/TimesNet repo (1,118 stars, MIT) now points to the Time-Series-Library, where the complete code and scripts live, so that is the maintained home. Like most TSLib models it is a trained-from-scratch point forecaster rather than a zero-shot foundation model, and the 2D machinery is heavier than simpler MLP forecasters.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (Wu et al., ICLR 2023): top rank on long-term and short-term forecasting, imputation, classification, and anomaly detection versus 15+ baselines (DLinear, FEDformer, Autoformer, Informer, Non-stationary Transformer). Community-reproduced within the TSLib benchmark suite.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: ICLR 2023 paper (arXiv 2210.02186), repo github.com/thuml/TimesNet (1,118 stars, MIT), maintained inside github.com/thuml/Time-Series-Library. Strengths: one backbone for five analysis tasks, strong published benchmark, permissive MIT license, maintained library home. Docked for: (1) accuracy claims are author-reported with limited independent blind replication; (2) heavier than simpler forecasters with similar point accuracy on many datasets; (3) no pretrained zero-shot variant. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · 2D temporal-variation modeling (ICLR 2023): author-reported top rank on five tasks (long/short-term forecasting, imputation, classification, anomaly detection) versus 15+ baselines.
- · Code maintained inside the Time-Series-Library; MIT licensed.
- · Claims are author-reported with limited independent blind replication; heavier than simpler forecasters; no pretrained zero-shot variant.