📈 Foundation models have arrived for time series. Zero-shot forecasting with no training required.
Just like GPT-4 can answer questions without fine-tuning, these models can forecast on datasets they’ve never seen:
1. ⚡ Chronos-2 (Amazon, 120M params) Encoder-only. Native support for known future covariates. Context length up to 8,192 steps.
2. 🦌 TiRex (NX-AI, 35M params) xLSTM-based. Zero-shot point and probabilistic forecasts across short and long horizons.
3. ⏱️ TimesFM 2.0 (Google, 500M params) Decoder-only for univariate forecasting. Supports 2,048+ point contexts without retraining.
4. 🔵 IBM Granite TTM R2 (IBM Research, from 1M params) The smallest option! Zero-shot and few-shot multivariate on a single GPU or even CPU.
5. 🌊 Moirai (Salesforce) Universal multi-frequency forecasting with calibrated confidence intervals.
Why does it matter? Previously you needed a separate model for each use case. Now you can get competitive forecasting with only historical data as input.
💡 Explanation in a nutshell#
A time series foundation model is like a universal forecasting expert: without ever seeing your company’s data, it can predict sales, energy demand, or any time series just from historical data. They eliminate the need to train models from scratch for each new problem.
More information at the link 👇
