Time series foundation models can be few-shot learners

Google Research · 9 months ago

Google Research developed TimesFM-ICF, an extension of their TimesFM foundation model that learns from a few examples at inference time using continued pre-training and learnable separator tokens rather than requiring task-specific fine-tuning. The model achieved 6.8% better accuracy than the base TimesFM and matched the performance of supervised fine-tuning without requiring complex user training. This allows businesses to deploy a single forecasting model across multiple tasks by providing just a few relevant examples instead of building specialized models for each forecasting application.