Code, model weights, and data will be released soon
Time series foundation models are finally taking off!
The previous articles explored 2 promising foundation forecasting models, TimeGPT and TimesFM.
This article will explore MOIRAI [1], a groundbreaking TS foundation model by Salesforce. MOIRAI is awesome in terms of performance — but more importantly, the authors have pledged to open-source the model and its training dataset!
This is mentioned in a tweet here by Caiming Xiong, VP of AI at Salesforce and one of the paper’s authors
The major contributions of this paper are the following:
- MOIRAI: A novel transformer-encoder architecture, functioning as a universal time-series forecasting model.
- LOTSA (Large Open Time Series Archive): The largest collection of open time series datasets with 27B observations across 9 domains.
- UNITS: An open-source library for training universal time-series models.
Moreover, this article discusses:
- How MOIRAI works and why it’s a powerful model.
- How MOIRAI performs compared to Google’s TimesFM
- MOIRAI benchmark results.
- Why MOIRAI will revolutionize the TS forecasting field.
Let’s get started.
I’ve launched AI Horizon Forecast, a newsletter focusing on time-series and innovative AI research. Subscribe here to broaden your horizons!
We described the challenges in detail here. To recap, these are:
- Difficulty finding public time-series data — for training a time-series foundation model.
- Time-series data are highly heterogeneous — unlike in NLP, where data have well-defined grammar and vocabulary.
- Time series can be multivariate — unlike in NLP, where input is one-dimensional.
- Time series have different granularities — daily, weekly, monthly, etc.