How much data do I need to train a model?

There is no universal minimum table count, entity count, or history window for good predictive performance. Start with relevant tables and enough labeled examples to evaluate the task on held-out data.

For temporal tasks, allow history for input features and subsequent observations for labels in each training, validation, and test period. Add data based on gaps in coverage and validation results.

See Predictive Queries for training windows and splits. The GPU Sizing guide covers hardware requirements, not minimum sample counts.