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

Selecting a synthetic data generator for pretraining time-series foundation models can double forecasting error, according to a new study that evaluated 11 generator families across two architectures [1]. The paper, posted to arXiv on June 6, 2026, found that under identical training budgets, the gap between the best-performing and worst-performing generator reached a factor of two in forecasting error [1]. The authors trained Chronos-T5-Mini and Moirai-Small from scratch using each generator family and observed that no single generator consistently outperformed the others across both model architectures [2]. Generator rankings proved unstable between the two architectures, meaning a generator that worked well for Chronos-T5-Mini did not necessarily perform well for Moirai-Small [2]. This finding undercuts the common assumption that a generator validated on one model family will transfer its utility to another [2]. Rather than attempting to solve the generator selection problem directly, the researchers tested a simple alternative: combining all 11 generator families into an equal-weight mixture [1]. That mixture matched or exceeded the performance of the best individual generator for both Chronos-T5-Mini and Moirai-Small [2]. When the mixture was further composed with real-world data, the resulting pretraining corpus produced the strongest forecasting results overall [2]. The work reframes synthetic pretraining as a corpus composition problem rather than a generator selection problem [1]. The authors recommend that composition choices be validated per model family instead of being assumed to transfer across architectures [2]. The study was conducted under the arXivLabs framework, which allows community collaborators to develop and share new features on the arXiv platform [1]. The findings carry implications for practitioners building time-series foundation models, where synthetic data generation has become a common strategy to supplement limited real-world datasets [1].