Google open‑sourced TimesFM model
Google released TimesFM, a time‑series foundation model trained on ~100 billion points that claims strong zero‑shot forecasting for traffic, weather and demand — a potential plug‑and‑play tool for financial time‑series experimentation. The model's zero‑shot promise makes it appealing for quick benchmarking against ARIMA or state‑space setups on market data. (x.com)
The TimesFM paper lists Abhimanyu Das, Weihao Kong, Rajat Sen and Yichen Zhou as primary authors and describes a decoder‑only transformer architecture presented at ICML 2024. Google Research published the TimesFM codebase and inference tooling under an Apache‑2.0 repo on GitHub with instructions, checkpoints and a recommendation that users have at least 32 GB of RAM for typical runs. Official checkpoints are hosted on Hugging Face, including the original timesfm-1.0-200m, the timesfm-2.0-500m checkpoint with 2048 context support, and the timesfm-2.5-200m checkpoint used in recent comparisons. TimesFM-2.0 expanded max context to 2,048 time points and experimented with multiple quantile heads, while TimesFM‑2.5 reduced the core parameter count to ~200 million, raised context support to 16,384 points and added an optional ~30M continuous quantile head for probabilistic forecasts. Benchmarking materials and third‑party coverage report TimesFM series models ranking at or near the top of GIFT‑Eval zero‑shot leaderboards (metrics cited include MASE and CRPS), with TimesFM‑2.5 reported as the zero‑shot leader on aggregated GIFT‑Eval scores. Google has operationalized TimesFM across its cloud stack: BigQuery ML documents a built‑in TimesFM univariate model and example notebooks show Vertex AI Model Garden integration for serving TimesFM checkpoints. A pip package and PyPI project page provide the community installer (timesfm on PyPI) and release notes that trace the open checkpoints and recommended inference flags, enabling direct local experimentation or cloud deployment with the published weights.
