TimesFM is a pretrained time-series foundation model created by Google Research for forecasting. The first open model checkpoint, timesfm-1.0-200m, focuses on univariate time series forecasting. It requires contiguous context and horizon of the same frequency, with the option for a frequency indicator. TimesFM does not support probabilistic forecasts but offers quantile heads experimentally. The model allows for context lengths up to 512 timepoints and any horizon length. To install and use TimesFM, follow the provided guidelines and refer to the extended benchmarks for results. The frequency indicator for forecasting can vary based on the time series data used.
https://github.com/google-research/timesfm