Mix, Don't Pick: Why Synthetic Corpus Composition Matters for Time Series Foundation Model Pretraining

In three linesStudy on synthetic corpus composition for time-series foundation model pretraining. Equal-weight mixture of 11 generator families outperforms best individual generators on Chronos-T5-Mini and Moirai-Small, reducing forecasting error gap up to 2×. Generator rankings vary across architectures.

Summary generated by Claude — human-verified