Google AI Unveils TimesFM-3: A 330M Parameter Zero-Shot Foundation Model for Multivariate Time Series Forecasting

Google Research has announced the release of TimesFM-3, a significant advancement in time series forecasting. This new foundation model boasts 330 million parameters and is specifically engineered for zero-shot multivariate forecasting, meaning it can predict multiple related time series simultaneously without requiring task-specific fine-tuning. This marks a substantial departure from its predecessors, as all previous TimesFM checkpoints, up to version 2.5, were strictly univariate. Univariate models could only forecast a single time series by analyzing its own historical data in isolation.

The development of TimesFM-3 addresses a critical limitation in many real-world forecasting scenarios. Unlike simplified academic problems, practical forecasting often involves intricate interdependencies. Google provides the example of ice cream sales, where predicting demand is not solely dependent on past sales figures but also influenced by a multitude of external factors. These can include the sales performance of related products, real-time foot traffic data, prevailing weather conditions, the impact of marketing promotions, and the timing of holidays. TimesFM-3 is designed to natively handle these complex relationships.

Pretrained on an unprecedented scale, TimesFM-3 was trained on over 1 trillion time points derived from both real-world and synthetic time series datasets. This extensive training enables it to process and understand complex patterns across numerous variables. The model's architecture builds upon a decoder-only transformer backbone, a common and powerful design in modern deep learning for sequence modeling. A key innovation in its processing involves grouping contiguous time points into 'patches' of 32 steps. These patches are then normalized per series, a crucial step that allows the model to effectively handle time series with vastly different scales and units, a common challenge in multivariate datasets.

TimesFM-3 demonstrates exceptional performance, achieving the top average rank among pretrained foundation models on several prominent forecasting benchmarks. These include GIFT-Eval, fev-bench, and the TIME leaderboard. Its success is measured across both point forecasting (predicting a single future value) and probabilistic forecasting (estimating a range of possible future values and their likelihoods), indicating its robustness and versatility.

Regarding deployment, Google has adopted a phased approach. The source code for the TimesFM repository is available under the permissive Apache-2.0 license, encouraging research and development. However, the actual TimesFM 3.0 model weights are distributed under a more restrictive timesfm-non-commercial-license-v1.0. This license explicitly prohibits commercial and production use. Consequently, while researchers and developers can readily benchmark and experiment with TimesFM-3 today, it cannot be integrated into production systems or offered as a commercial forecasting API. This licensing strategy allows for broad accessibility for research purposes while maintaining control over its commercial application.