MOMENT

MOMENT (Goswami et al., ICML 2024) is the first fully open time-series foundation model family built end to end for open science. The team compiled the Time-Series Pile, a large diverse public collection, pre-trained T5-based models (small, base, large up to 385M parameters) with masked patch reconstruction, and released code, weights, training logs, and the dataset. One backbone then serves forecasting, classification, anomaly detection, imputation, and embedding extraction through a clean pipeline: MOMENTPipeline.from_pretrained on Hugging Face (AutonLab/MOMENT-1-large) or pip install momentfm.

The authors are refreshingly honest in their own results: pre-training helps most on anomaly detection and classification, while simple statistical and shallow methods like ARIMA and N-BEATS still win on several forecasting settings. Zero-shot forecasting works with no fine-tuning; linear probing covers the rest. MOMENT is the go-to when you want a single open backbone that also handles non-forecasting tasks, at the cost of not leading any single forecasting leaderboard.

Author-reported (Goswami et al., ICML 2024): better than statistical imputation baselines zero-shot; second-best anomaly F1 zero-shot, best with linear probing; better than 11/16 compared methods on classification; competitive but not leading on long-horizon forecasting. The paper's own ablations show ARIMA and N-BEATS beating it on several forecasting tasks.