Granite Time Series PatchTST-FM-r2 is a Large Language Models (LLMs) tool. Open-source foundation model for zero-shot time series forecasting with 385M parameters. Key features include Zero-Shot Forecasting, Probabilistic Predictions, and Long Context Support. Best for data scientists and analysts, financial advisors and analysts and scientists and researchers.
About Granite Time Series PatchTST-FM-r2
Key Features
<strong>Zero-Shot Forecasting.</strong> The model generates accurate forecasts on new data without requiring fine-tuning or task-specific training, making it ready to use out of the box.
<strong>Probabilistic Predictions.</strong> It predicts 99 quantiles to provide full uncertainty distributions rather than single point estimates, helping users understand forecast confidence and risk.
<strong>Long Context Support.</strong> The model handles context lengths up to 8,192 time steps, allowing it to process extensive historical data for better long-term predictions.
<strong>Missing Value Imputation.</strong> Built-in support for automatically handling and imputing missing values in time series data without requiring external preprocessing.
<strong>Patch-Based Architecture.</strong> Time series are segmented into overlapping patches that capture local patterns while reducing computational complexity compared to traditional transformers.
<strong>Open Source and Permissive Licensing.</strong> Released under Apache 2.0 and OpenMDW 1.0 licenses, the model is free for both research and commercial use with full code and weights available on Hugging Face.
Frequently Asked Questions
Granite Time Series PatchTST-FM-r2 is used for time series forecasting across various domains including finance, energy demand prediction, inventory planning, and any application requiring predictions from historical sequential data. It works with multivariate time series and provides both point forecasts and uncertainty estimates.
Yes, Granite Time Series PatchTST-FM-r2 is completely free and open source. It's released under Apache 2.0 and OpenMDW 1.0 licenses, making it available for both research and commercial use without restrictions or fees.
No, Granite Time Series PatchTST-FM-r2 is designed for zero-shot forecasting, meaning it can generate predictions on your data immediately without training. However, you can optionally fine-tune it on your specific data if you want to improve accuracy further.
Granite Time Series PatchTST-FM-r2 ranks second among replicable zero-shot models on the GIFT-Eval benchmark for both CRPS and MASE metrics. With 385 million parameters, it balances performance and efficiency while supporting longer context windows than many competing models.
