Hands-On with Moirai: A Foundation Forecasting Model by Salesforce | by Marco Peixeiro

Discover the architecture and inner workings of Moirai and apply it in a forecasting project using Python

We have entered an era where large foundation models are now common. Foundation models have revolutionized many fields, like computer vision and natural language processing, with models and applications that can generate text, images and videos.

The field of time series forecasting has not been impervious to this movement, with many foundation models appearing for forecasting. This marks an important paradigm shift, as we can now generate zero-shot predictions of time series data and avoid the costs and development time of training data-specific models.

In October 2023, TimeGPT-1 was published, marking the appearance of one of the first foundation forecasting models. Then, in February 2024, Lag-Llama was released, which was quickly followed by Chronos in March 2024.

In May 2024, a new open-source foundation forecasting model was released: Moirai. In the paper Unified Training of Universal Time Series Forecasting Transformers, researchers from Salesforce propose a foundation model capable of probabilistic zero-shot forecasting, while also supporting exogenous features.