Moirai is a time-series foundation model from Salesforce AI Research, introduced in “Unified Training of Universal Time Series Forecasting Transformers,” posted to arXiv in February 2024. The name refers to the Fates of Greek myth, fitting for a model meant to forecast any series.
Moirai aims to be a single model that forecasts time series of any frequency, any number of variables, and any domain, without per-dataset training. To make that possible the authors built LOTSA, the Large-scale Open Time Series Archive, with more than 27 billion observations spanning nine domains, and designed the transformer to handle the awkward variety of real time series: different sampling rates, multiple correlated variables, and varied patterns. The paper reports that as a zero-shot forecaster Moirai matches or beats models trained specifically on each target dataset.
Alongside TimeGPT, Chronos, and TimesFM, Moirai is part of the 2024 wave of foundation models that brought the pre-train-once, forecast-anything paradigm to time series.
Why business readers should care: Moirai targets the messy reality of enterprise data, many series at different frequencies and dimensions, with one model that needs no bespoke training for each new dataset.