Google unveils timesfm-3: a multivariate time series forecasting model with 330m parameters

Google Research has introduced TimesFM-3, a zero-shot foundation model designed for multivariate time series forecasting. The model, with 330 million parameters, is pretrained on over 1 trillion time points and can handle multiple targets, past covariates, and past-future covariates without task-specific fine-tuning. It outperforms other models on GIFT-Eval, fev-bench, and the TIME leaderboard, achieving top rankings in both point and probabilistic metrics. TimesFM-3 uses a decoder-only transformer architecture with contiguous patch masking, allowing it to process data efficiently. It groups contiguous time points into patches of 32 steps and applies normalization to manage varying scales. The model also employs a lookahead trick to incorporate future data, enhancing its predictive capabilities. Google highlights its application in forecasting ice cream sales, considering factors like weather, promotions, and holidays. The model's weights are available under a non-commercial license, restricting their use to non-production environments. Google encourages benchmarking but cautions against deploying it in production APIs. The release includes technical details, model weights on Hugging Face, and a GitHub repository for further exploration.