ContestTrade

ContestTrade runs a two-stage pipeline. Parallel Data Analysis Agents refine raw multi-source data into structured “textual factors”, and an internal contest builds a factor portfolio from them; Research Agents, each conditioned on a distinct “Trading Belief”, then produce proposals that a second contest round synthesises into one allocation. Beliefs are plain text a user edits — contest_trade/config/belief_list.json holds a JSON array of belief strings, and each belief emits at most five signals. There is no execution layer: the output is signals and Markdown research reports written to contest_trade/agents_workspace/results, not orders. A published Docker image (finstep/contesttrade:v2.0) is an alternative to the clone-and-pip path.

The README states the project is intended for academic and educational purposes only, warns explicitly about model hallucination and data inaccuracy, and publishes no performance figures of its own. The backtest claims live only in the accompanying preprint, arXiv:2508.00554, submitted August 2025 and revised through a fourth version in July 2026; the abstract page names no conference or journal, so the paper is unrefereed and its reported improvement over benchmarks on post-2024 Chinese A-shares is author-reported. Note also that the arXiv author list and the README’s own BibTeX disagree on order and membership — the arXiv page is the authoritative one, and is what this entry follows.

The code has been still since 22 December 2025, roughly eight months as of this entry, while the paper kept being revised; issues continue to arrive without visible maintainer replies. Finished roadmap items are the V1.1 data-provider refactor and CLI rework and V2.0 US-stock market access with richer factor and signal sources; Hong Kong stocks, a visual backtesting interface and agent scale-up remain unchecked. The English documentation is README_en.md — the repository’s default README.md is Chinese — and the Apache LICENSE appendix was left as the unfilled template, so it names no copyright holder, though the grant itself is unaffected.