TimesFM-3: A zero-shot foundation model for multivariate forecasting (research.google)

Google Research has unveiled TimesFM-3, an advanced foundation model designed for multivariate time series forecasting, which markedly improves forecast accuracy in a single forward pass. This new model significantly advances the capabilities previously established by TimesFM-2.5, which was limited to univariate forecasting. Now, TimesFM-3 can synergistically predict multiple coevolving time series using a pre-trained architecture with 330 million parameters, built on a massive dataset encompassing over 1 trillion time points. Notably, it employs an innovative Contiguous Patch Masking strategy, allowing it to fill in forecasts for an entire horizon at once, drastically reducing both latency and error propagation seen in previous models. The introduction of TimesFM-3 holds tremendous significance for the AI/ML community, particularly in industries reliant on accurate forecasting such as retail, finance, and healthcare. By incorporating external covariates, such as promotions or seasonal trends, the model can anticipate significant impacts on sales, illustrated by its ability to predict a 20% sales increase during promotional events far more accurately than traditional methods. TimesFM-3 has already set new benchmarks for forecasting accuracy across three major public datasets, and it is now accessible on platforms like GitHub and Hugging Face, fostering broader adoption and experimentation in multivariate time series forecasting, all without requiring extensive machine learning expertise.