Google Releases TimesFM-3 Time Series Forecasting AI, Capable of Handling Multiple Data Streams Without Additional Training

This model supports multivariate forecasting, allowing it to simultaneously process multiple data streams that change over time—such as sales, foot traffic, and weather—to predict future values.

TimesFM-3 features approximately 330 million parameters and has been pre-trained on over 1 trillion data points. Because it has learned a vast array of patterns, it is capable of "zero-shot forecasting," meaning it does not require additional training for new datasets.

Improvements have also been made to forecasting speed. Rather than predicting values sequentially as previously done, the model introduces a mechanism that predicts values for the target period in a single batch process. According to Google, this reduces processing time and suppresses the accumulation of errors.

Google announced that in evaluations using three types of public benchmarks, the model recorded the top average rank in both "point forecasting" and "probabilistic forecasting" among the pre-trained models compared.

It is currently available via GitHub and Hugging Face. While the source code is under the Apache License 2.0, the pre-trained models are subject to a non-commercial license, and commercial use is not permitted.

Google stated that it plans to integrate TimesFM-3 into its data analysis service, "BigQuery," within a few weeks.