Google's GPT-3 Moment for Time Series & Forecasting (TimesFM 3.0)
Thanks to Neon for sponsoring this video. Try Neon here : https://neon.plug.dev/Sw0JRgk Google's TimesFM 3.0 forecasts time series without any training on your data. Hand it the history, and it continues the chart with a range around every day. In this video, I show how it works (patches, tokens, "autocomplete for charts"), what changed from 2.5 to 3.0, and I test both versions on real New York subway ridership and on a simulated business running on Neon. Then I give the exact planning decision to a Gemini agent, and the two versions give opposite answers. This video is sponsored by Neon, the backend platform for apps and agents: Postgres, auth, storage and functions in one place. Every what-if scenario in this video runs on its own Neon branch, a copy-on-write clone of the database that's ready in about a second. Links: - Google Research blog, TimesFM-3: https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/ - TimesFM on GitHub: https://github.com/google-research/timesfm - TimesFM 3.0 weights (non-commercial license): https://huggingface.co/google/timesfm-3.0-pytorch - Subway data, MTA Daily Ridership (data.ny.gov): https://data.ny.gov/Transportation/MTA-Daily-Ridership-and-Traffic-Beginning-2020/sayj-mze2 - My TabFM video (foundation models for tabular data): https://youtu.be/5YJGDElbCDY Website: https://engineerprompt.ai/ RAG Beyond Basics Course: https://prompt-s-site.thinkific.com/courses/rag Let's Connect: 🦾 Discord: https://discord.com/invite/t4eYQRUcXB ☕ Buy me a Coffee: https://ko-fi.com/promptengineering |🔴 Patreon: https://www.patreon.com/PromptEngineering 💼Consulting: https://calendly.com/engineerprompt/consulting-call 📧 Business Contact: engineerprompt@gmail.com Become Member: http://tinyurl.com/y5h28s6h 💻 Pre-configured localGPT VM: https://bit.ly/localGPT (use Code: PromptEngineering for 50% off). Signup for Newsletter, localgpt: https://tally.so/r/3y9bb0 Timestamps: 0:00 Can AI predict the future? 1:40 What TimesFM is (and isn't) 2:18 Why one model can forecast any series 2:47 How it works: patches, tokens, autocomplete for charts 3:43 What's new in TimesFM 3.0 vs 2.5 4:54 Test 1: NYC subway ridership 5:36 Labor Day: why context matters 6:09 Results: 2.5 vs 3.0 on real data 7:03 Sponsor: Neon 7:52 Lumen: a simulated business on Neon 8:23 Backtest from a Neon branch 8:48 Nightly forecasts and a quarter later 9:08 The planning question 9:57 Handing it to a Gemini agent 10:24 Where the two versions disagree 11:39 4 limitations to know 13:11 The verdict #TimesFM #TimeSeries #Forecasting #FoundationModels #Neon
