Reproduce “QUALS: Corpus Equilibrium for Universal Forecasting via Pattern Quantization and Learnability Synchronization” on OpenResearch

About this paper

Yujie Li, Zezhi Shao, Chengqing Yu, Yisong Fu, Weijie Zhu, Yifan Du, Jilin Hu, Bin Yang and 2 more

QUALS presents a data-centric framework for universal time-series forecasting that organizes pre-training data by learned temporal patterns and then adjusts sampling according to pattern difficulty. Across several forecasting model families and benchmarks, it improves zero-shot accuracy while reducing the amount of training data required, including a reported reduction from 419.43 billion to 26.20 billion tokens for TimeMoE-Base at lower average MSE.

Key insights

• Effective corpus design requires two forms of balance: coverage across temporal structures and synchronization of training exposure with pattern learnability.
• Learned multi-scale representations outperform purely handcrafted statistical descriptors for organizing heterogeneous time-series corpora, while the number of pattern bins must balance pattern separation against data sparsity.
• The framework is architecture-compatible and performs corpus selection and sampling offline, avoiding repeated full-model retraining during policy optimization.

Summary by alphaXiv.