方法对比

	Moirai：通用时间序列预测Transformer	PatchTST	TimesFM：面向时间序列预测的仅解码器基础模型
领域	深度学习	深度学习	深度学习
方法族	Machine learning	Machine learning	Machine learning
起源年份	2024	2023	2024
提出者	Gerald Woo et al. (Salesforce)	Nie, Y. et al.	Abhimanyu Das et al. (Google)
类型	Foundation model for zero-shot time-series forecasting	Transformer for time series forecasting	Pre-trained decoder-only transformer for zero-shot time-series forecasting
开创性文献	Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., & Sahoo, D. (2024). Unified training of universal time series forecasting transformers. ICML.	Nie, Y., Nguyen, N. H., Sinthong, P. & Kalagnanam, J. (2023). A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. ICLR.	Das, A., Kong, W., Sen, R., & Zhou, Y. (2024). A decoder-only foundation model for time-series forecasting. ICML.
别名	Unified Time-Series Transformer, Universal Forecasting Transformer, MOIRAI, Evrensel Zaman Serisi Tahmin Transformatörü	PatchTST — Yama Tabanlı Zaman Serisi Transformer, patch-based time series transformer, channel-independent transformer	Time-series Foundation Model, Google TimesFM, TimesFM forecaster, Zaman Serisi Temel Modeli
相关	3	3	3
摘要	Moirai is a foundation model for universal time-series forecasting introduced by Gerald Woo and colleagues at Salesforce Research in 2024 and presented at ICML. The core idea is to pre-train a single large Transformer on an exceptionally diverse corpus of time-series data (LOTSA) spanning many domains and frequencies, enabling zero-shot and few-shot forecasting on unseen datasets without task-specific retraining. Moirai employs patch-based tokenization, any-variate attention, and a mixture-of-distributions output head to handle variable frequencies, multiple variates, and probabilistic prediction in a unified architecture.	PatchTST is a patch-based Transformer architecture for time series forecasting, introduced by Nie and colleagues in 2023, that cuts each series into overlapping patches treated as tokens and processes channels independently. It balances computational efficiency with strong accuracy on long-horizon forecasting.	TimesFM is a pre-trained foundation model for univariate time-series forecasting introduced by Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou from Google in 2024. The model adopts a decoder-only transformer architecture, similar in spirit to large language models, and is trained on a large corpus of real-world and synthetic time-series data. Its central innovation is the ability to perform accurate zero-shot forecasting across diverse domains without task-specific fine-tuning.