Google Research releases TimesFM 2.5 for zero-shot time-series forecasting

Google Research has released TimesFM 2.5, an open-source foundation model for time-series forecasting. This model, with 200 million parameters and a context window of up to 16,384 points, can predict future trends without requiring task-specific training data, operating in a zero-shot capacity. It demonstrates improved accuracy over traditional methods like Auto ARIMA, showing a 15-25% reduction in Mean Absolute Error on benchmarks, though its performance on certain stable time series may be comparable to classic approaches. While the model weights are available on Hugging Face, some of Google's associated cloud services are inaccessible from Russia. AI

IMPACT This open-source model could lower the barrier for businesses to adopt advanced time-series forecasting, potentially improving inventory management and resource planning.

RANK_REASON Frontier-lab model release with system card. [lever_c_demoted from frontier_release: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• arXiv
• Auto ARIMA
• BigQuery
• GPT
• GIFT-Eval
• Google Research
• Google Trends
• Grid Dynamics
• Hugging Face
• ICML 2024
• Rajat Sen
• TimesFM 2.5
• Wikipedia
• Yichen Zhou

• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →