google-research

timesfm

AI#Time Series#Forecasting#Foundation Model#PyTorch#JAX

// summary

TimesFM is a decoder-only foundation model developed by Google Research specifically for time-series forecasting tasks. The latest 2.5 version features a 200M parameter architecture that supports up to 16k context length and continuous quantile forecasting. The repository provides comprehensive tools for inference, fine-tuning with LoRA, and integration with agentic workflows.

// technical analysis

TimesFM is a decoder-only foundation model developed by Google Research specifically designed for time-series forecasting. By leveraging a large-scale pretrained architecture, it addresses the challenge of building robust, general-purpose forecasting models that can handle diverse temporal data without requiring extensive task-specific training. The project balances performance and efficiency by offering a 200M parameter model that supports significantly longer context lengths and optional quantile heads, providing a versatile tool for both enterprise-level applications and custom research workflows.

// key highlights

Supports an extended context length of up to 16k, allowing for the analysis of much longer historical time-series data.

Features a 200M parameter architecture that improves efficiency while maintaining high forecasting performance.

Includes an optional 30M parameter quantile head to enable continuous quantile forecasting for up to 1k horizons.

Provides flexible backend support, allowing users to choose between PyTorch and Flax for optimized inference across different hardware.

Enables fine-tuning capabilities through integration with HuggingFace Transformers and PEFT (LoRA) for domain-specific adaptation.

Incorporates covariate support via XReg, allowing the model to account for external variables in its forecasting process.

// use cases

High-performance time-series forecasting with support for continuous quantile outputs

Fine-tuning the foundation model using HuggingFace Transformers and PEFT (LoRA)

Integration into agentic workflows and enterprise-level applications via BigQuery ML and Vertex AI

// getting started

To begin, clone the repository and set up a virtual environment using uv. Install the package in editable mode with your preferred backend—either torch, flax, or xreg—and ensure the corresponding deep learning framework is installed for your hardware. You can then load the pretrained model using the provided Python API and configure the forecast parameters to generate predictions on your time-series data.