Reproduce “Loss-Guided Pretraining Data Selection for Time-Series Foundation Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yike Li, Shaoxu Song, Jianmin Wang
The authors propose selecting pretraining windows for time-series foundation models by scoring each context–future window with a frozen reference forecaster, ranking scores within each source dataset, and retaining a middle band. At 80% retention, this selection improves MASE and CRPS over full-data and size-matched random training for Chronos-Bolt Base, Moirai-Base, and TimesFM 2.5, but still scores every candidate offline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.