Reproduce “Fracast-0: Fractal Weight Sharing for a Time Series Foundation Model with Only 85K Parameters” 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

Tianxiang Zhan, Huanyao Zhang, Yuanpeng He

Fracast-0 forecasts unfamiliar time series by reusing shared local operators across temporal scales, producing a median forecast and nine quantiles. Across 97 GIFT-Eval configurations spanning datasets, frequencies, and horizons, its Seasonal-Naive-normalized geometric-mean MASE and WQL are 0.808 and 0.564, with lower values better. At 85,001 parameters, it is a frontier point, but not strict zero-shot evidence because training overlaps test-data families.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.