TimesFM 1.0

TimesFM 1.0 is the first open checkpoint of Google Research's foundation model for time-series forecasting. The paper behind it appeared in October 2023 and was accepted at ICML 2024; the weights were published on 8 May 2024, when Google updated the research blog post to say the model was now on Hugging Face and GitHub. The idea was borrowed from language models: instead of fitting a separate statistical model to every series, pretrain one model on an enormous corpus of series and let it forecast new ones zero-shot. The corpus was 100 billion real-world time-points, the majority of it search interest data from Google Trends and pageviews from Wikipedia — two sources that, as the authors put it, mirror the trends and seasonality found in many other real-world series. The checkpoint is deliberately small: 200 million parameters, a decoder-only transformer over patches, 20 layers of dimension 1,280, history read in patches of 32 points and predicted in patches of 128. That asymmetry is the design decision the paper argues for most strongly — a long output patch means fewer autoregressive steps to reach a distant horizon, and therefore less accumulated error. The limits of the first generation are equally clear: a maximum context of 512 points, a frequency indicator that the user has to supply, and point forecasts only, since the quantile heads shipped with the model were never calibrated. Google's published comparisons are charts rather than tables, so this profile carries no figures from them — but their direction is unambiguous. On the Monash Forecasting Archive, measured as the geometric mean of scaled MAE, the zero-shot model beats most supervised approaches, including recent deep-learning ones, and also beats llmtime with GPT-3.5 while being orders of magnitude smaller. On the ETT long-horizon sets at 96 and 192 points ahead it matches PatchTST trained directly on those datasets. Three generations later the traffic has moved on — around three thousand downloads a month across both mirrors, against 86,000 for TimesFM 2.0 — yet this repository still holds the most likes in the family, 840, more than the current TimesFM 3.0. The weights are Apache 2.0 and remain available.

▸Videos

No videos yet.

▸Reviews