TimesFM is a time-series foundation model from Google Research, described in “A decoder-only foundation model for time-series forecasting” by Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou, posted to arXiv in October 2023. It applies the decoder-only design that powers large language models to numeric forecasting.
The model uses a patched-decoder attention architecture: it breaks a time series into patches (short contiguous chunks of values), treats each patch like a token, and predicts forward in the autoregressive, decoder-only style of a GPT-type model. It was pre-trained on a large and diverse corpus of time-series data, which lets it deliver strong zero-shot performance on a variety of public datasets it was not trained on, adapting across different forecast horizons and time granularities. Notably, TimesFM is relatively compact compared with text foundation models yet still generalizes well, which makes it practical to deploy.
TimesFM, released around the same time as TimeGPT and slightly before Chronos and Moirai, was part of the cluster of 2023-2024 efforts establishing foundation models as a new default approach to forecasting.
Why business readers should care: TimesFM offers Google-scale pre-training so a company can get accurate forecasts on its own series without training a model, lowering the bar to good forecasting.