Introducing TimesFM-3
Google Research has announced TimesFM-3, a new time series foundation model designed for multivariate forecasting. The model’s core capability is its ability to accurately predict multiple time series simultaneously, a significant advancement over previous models that were limited to univariate forecasting.
Model Specifications
The TimesFM-3 model utilizes a decoder-only transformer architecture with 330 million parameters. It was pre-trained on a corpus of over 1 trillion time points, encompassing both real-world and synthetic data. The model’s architecture incorporates causal temporal attention and full variate attention mechanisms, enabling it to capture complex dependencies between time series.
Forecasting Capabilities
TimesFM-3 supports multiple targets, allowing for the simultaneous forecasting of related time series. It can incorporate past covariates, such as historical foot traffic, and past-future covariates, like planned promotional campaigns or weather forecasts. The model generates 9 quantiles for each target time series at every forecasting horizon, providing a comprehensive probabilistic view of the forecast uncertainty. The model employs Contiguous Patch Masking to generate the entire forecasting horizon in a single forward pass.
Performance and Use Cases
Evaluations demonstrate TimesFM-3’s superior performance compared to other forecasting models across major benchmarks. The model’s zero-shot generalization capabilities allow it to handle complex multivariate scenarios without requiring task-specific fine-tuning. Applications include retail, finance, observability, manufacturing, healthcare, and natural sciences, where accurate multivariate forecasting is crucial for decision-making.
Source: https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/