PatchTST

PatchTST (Nie, Nguyen, Sinthong, Kalagnanam, ICLR 2023) is the paper that made patching the default transformer trick for time series. It borrows the idea straight from vision transformers: split each univariate series into fixed-length patches, feed the patches as tokens, and process each channel independently with shared weights. Patching keeps local semantics in the embedding, quadratically cuts attention memory, and lets the model attend to much longer history. A second head does self-supervised masked-patch pretraining that transfers across datasets.

The authors report PatchTST/64 reducing MSE by 21.0% and MAE by 16.7% versus the best transformer baselines of the time. The official yuqinie98/PatchTST repo (2,726 stars, Apache 2.0) is now complemented by implementations in Hugging Face transformers, GluonTS, NeuralForecast, and tsai, which is the practical on-ramp. PatchTST is a point forecaster with no native probabilistic output, and as a trained-from-scratch model it does not do zero-shot like the newer foundation models.

Author-reported (Nie et al., ICLR 2023): PatchTST/64 gives 21.0% MSE and 16.7% MAE reduction over the best transformer baselines (DLinear, FEDformer, Autoformer, Informer, Pyraformer, LogTrans) on long-term benchmarks; self-supervised variant beats supervised baselines on transfer. Widely reproduced by the community with consistent results.