Reproduce “KiT: A Foundation Model for Financial Time-Series Forecasting using DiffusionTransformers” 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
Boyu Zhang, Haorui Li
KiT generates candlestick paths jointly rather than forecasting bars one at a time. It uses an invertible five-feature representation and flow matching to produce OHLCV paths, with sampling requiring Euler integration. On the 2026 test window, mean return RankIC is 0.0571 versus 0.0521 for Kronos-base-FT; results vary by market, and that checkpoint was selected using test scores.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.