Tiny-Time-Mixers R2 (TTM): More Accurate Predictions with Exogenous Feature Mixing - Complete Tutorial

IBM’s open-source foundation model for time-series just got even better!

Of all foundation time-series models, IBM’s TTM has a unique feature:

It is the first non-Transformer-based foundation TS model.

Tiny-Time-Mixers (TTM)[1] is a breakthrough in time-series forecasting, delivering high-accuracy predictions with zero-shot capability or minimal fine-tuning.

We previously explored its mechanics in depth here. Since then, the model has received major upgrades—including the R2.1 variant, which now handles daily and weekly seasonalities.

TTM is a lightweight, open-source model that continues to evolve. Its latest version (the R2 series) secured 1st place on the Gift-Eval Benchmark for point forecasting.

This article covers how Tiny-Time-Mixers works and how to leverage its advanced features, such as feature-mixing and exogenous infusion—focusing on the latest R2 version.

Let’s get started!

✅ Find the Tiny-Time-Mixers notebook in the AI Projects folder (Project 16)

🎁 Subscribe to AI Horizon Forecast by next week to receive a -20% discount. You'll gain access to all my AI Capstone Projects, including this interesting project on using Tiny-Time-Mixers!

Enter Tiny-Time-Mixers

The first key innovation of TTM is its lightweight, pretrained design, eliminating the need for costly self-attention. Moreover, TTM offers: