- Published on
A 330M Parameter Zero-Shot Foundation Model For Multivariate Time Series Forecasting
- Authors
- Name
- DeepHorus
Google Research has unveiled TimesFM-3, a significant advancement in time series forecasting with a 330 million parameter foundation model designed for multivariate analysis. This release marks a departure from previous TimesFM versions, which were limited to univariate forecasting, focusing on a single series from its own history. The new model is pretrained natively for multivariate forecasting on more than 1 trillion time points and can handle multiple targets, past covariates, and past-future covariates without any task-specific fine-tuning.
The development addresses a common real-world scenario where multiple interrelated factors influence a target variable. For example, ice cream sales are affected by related product sales, foot traffic, weather, promotions, and holidays. TimesFM-3 can process all these inputs together in a single forward pass, providing both point and quantile outputs for each target, a capability essential for probabilistic forecasting.
Technically, TimesFM-3 uses a decoder-only transformer backbone. It groups contiguous time points into patches of 32 steps, normalizing per series to handle different scales. The architecture features two alternating attention mechanisms: causal temporal attention that runs horizontally to prevent data leakage, and full variate attention that runs vertically to learn correlations across all series at a given time step.
A key innovation is the use of Contiguous Patch Masking, a training strategy that allows the model to output the entire forecast horizon simultaneously. This approach replaces the sequential patch decoding of earlier versions, reducing latency, compute cost, and compounding error. For each horizon step, the model provides 9 quantiles, from the 10th to the 90th percentile.
In benchmarks, TimesFM-3 has demonstrated leading performance. It achieved the top average rank among pretrained foundation models on the GIFT-Eval, fev-bench, and TIME leaderboards. The model secured rank #1 overall on fev-bench across 100 tasks, rank #1 overall on TIME across 98 tasks, and rank #1 among foundation models on GIFT-Eval.
The release includes both the model and code, but with important usage restrictions. While the code repository is Apache-2.0 licensed, the TimesFM 3.0 weights are available under a non-commercial license. This means the model can be used for benchmarking and research, but it cannot be deployed in production systems or commercial APIs at this time.
Looking ahead, TimesFM-3 sets a new standard for zero-shot multivariate forecasting capabilities. Its ability to model cross-series dependencies in a single pass could transform applications requiring complex, multi-input predictions. However, commercial deployment will require future licensing changes, leaving its production use contingent on Google's ongoing support and business strategy for the model.
