MOIRAI-2: Faster, Lighter, More Accurate Time-Series forecasting
The flagship foundation forecasting model just got an upgrade!
MOIRAI was among the first models to spark the foundation-model revolution in time-series forecasting.
Beyond the model itself, the authors released valuable contributions for the field: the LOTSA pretraining dataset, the GIFT-Eval benchmark, and MOIRAI’s fine-tuning scripts.
MOIRAI stood out early on thanks to its support for external covariates, flexible context and prediction lengths, and probabilistic forecasting.
Now the authors are back with an improved version: MOIRAI-2. While the paper isn’t out yet, the results speak for themselves—MOIRAI-2 ranks #1 on the GIFT-Eval benchmark by MASE, among models that aren’t affected by test data leakage.
In this article, we’ll put MOIRAI-2 to the test against MOIRAI-1 and strong statistical baselines such as AutoETS and Dynamic Optimized Theta.
Let’s get started!
✅ Find the MOIRAI-2 tutorial with my benchmark in the AI Projects Folder (Project 22) — plus more cool projects inside!
Enter MOIRAI-2
Here’s what we know so far about Moirai 2.0: