google-research / timesfm

2024-04-29 · Source: Github Trending: All languages · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, quick

Summary

Google Research has released TimesFM 2.5, a decoder-only time-series foundation model for forecasting, available via Hugging Face and integrated into Google BigQuery. This latest version significantly reduces parameters from 500M to 200M while expanding context length support to 16k, up from 2048. TimesFM 2.5 also introduces continuous quantile forecasting for horizons up to 1k via an optional 30M quantile head and removes the need for a frequency indicator. The model's inference API has been upgraded, with future plans to support a Flax version for faster inference and reintroduce covariate support. Installation involves cloning the GitHub repository and using `uv` to set up a virtual environment with either PyTorch or Flax dependencies.

Key takeaway

For AI Engineers and Research Scientists building forecasting systems, TimesFM 2.5 offers a compelling upgrade. Its reduced parameter count (200M) and expanded context window (16k) mean more efficient and accurate predictions, especially for complex, long-range time series. You should consider migrating to TimesFM 2.5 to benefit from improved performance and new features like continuous quantile forecasting, potentially simplifying your model architecture and deployment.

Key insights

TimesFM 2.5 is a more efficient, capable time-series foundation model with reduced parameters and extended context.

Principles

• Smaller models can achieve greater capabilities.
• Foundation models enhance time-series forecasting.

Method

Install TimesFM 2.5 via `uv` with PyTorch or Flax, then load and compile the model with a `ForecastConfig` to generate point and quantile forecasts.

In practice

• Use TimesFM 2.5 for long-context time-series forecasting.
• Leverage the quantile head for probabilistic forecasts.
• Integrate with BigQuery for cloud-based forecasting.

Topics

• TimesFM
• Time-series Forecasting
• Foundation Models
• Google Research
• Quantile Forecasting

Code references

• google-research/timesfm
• google-research/timesfm

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• #356 The Forecast for Time Series Forecasts with Rami Krispin, Senior Manager of Data Science at Apple — Time series foundation models enable scalable forecasting for massive datasets, moving beyond traditional methods.
• The digital quant: instant portfolio optimization with JointFM — JointFM is a zero-shot foundation model predicting full multivariate distributions for instant, domain-agnostic time-series forecasting.
• 5 Useful Python Scripts for Time Series Analysis — Five Python scripts simplify common time series analysis tasks, from data preparation to forecasting and comparison.
• Representing Time Series as Structured Programs for LLM Reasoning — Representing time series as structured symbolic programs enables LLMs to reason effectively without fine-tuning.
• Why Do Time Series Models Need Long Context Windows? — Long context windows in time series forecasting primarily reduce uncertainty in identifying the data-generating process, crucial for optimal predictions.
• Assessing the Performance-Efficiency Trade-off of Foundation Models in Probabilistic Electricity Price Forecasting — TSFMs generally outperform task-specific models in PEPF, but well-configured conventional models remain highly competitive.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Github Trending: All languages.