[DigitalToday reporter Jae-won Choi (최재원)] Google unveiled TimesFM-3, an AI model that forecasts future figures by analysing multiple time-series datasets at once, including sales, customer numbers and weather. Its key features are zero-shot prediction, which can be applied immediately to new data without additional training, and multivariate analysis.

On Sept. 1 (local time), Japanese media outlet Gigazine reported that TimesFM-3 is the latest version of the time-series forecasting foundation model Google has been developing since 2024. While TimesFM-2.5 mainly forecast the future based on a single time-series record, TimesFM-3 processes multiple target variables along with past auxiliary data and external variables such as schedules and weather that are already known for the future.

For example, when forecasting ice cream sales, users can enter not only past sales volume but also customer numbers, upcoming weather and promotion schedules at the same time. In an example released by Google, it learned the relationship between past promotions and sales and projected about a 20 percent increase in sales on each future event day. Google said the example is intended to explain how the model works, not actual corporate performance.

The model has about 330 million parameters. It was pre-trained on more than 1 trillion time-series data points combining real and synthetic data. It can analyse relationships across multiple time series without additional task-specific fine-tuning, and it adopts a non-autoregressive approach that computes in a single forward pass without sequentially generating the entire future segment. Google said this reduced processing latency and the accumulation of errors in repeated forecasting steps.

Three public benchmarks were used for performance evaluation: GIFT-Eval, FEV-Bench and TIME. Google said TimesFM-3 ranked highest on average for point forecasts and probabilistic forecasts across all three benchmarks among the compared pre-trained time-series foundation models. In univariate mode, it matched or outperformed competing models, and performance improved further in multivariate mode using multiple variables together.

TimesFM-3 is currently available for download on GitHub and Hugging Face. The source code is released under the Apache License 2.0, but the pre-trained model weights are under a separate non-commercial licence, meaning it cannot be used directly for commercial services or real operating environments. Google plans to integrate TimesFM-3 into BigQuery within weeks. From a business perspective, its potential use is drawing attention because it can forecast interlinked data such as sales, demand, inventory and traffic without separate model training.