English README | 中文故事版 / Codex Series 01
OpenAI Codex + GPT-5.5 assisted A-share quantitative finance research toolkit for after-close stock selection, backtesting, paper replay, and manual trading review.
Case study: I gave an A-share account to Codex and doubled it in one month
BaiQuant turns a daily A-share research routine into reproducible code: pull local data, rank candidates, review risk, record real fills, and learn from the next session.
| Built with | OpenAI Codex + GPT-5.5 |
| Research style | Harness-style loop: idea -> rule -> code -> test -> backtest -> review |
| Market focus | China A-share equities |
| Public strategy | steady-20d |
| Operating mode | after-close plan, next-session manual execution |
| Data stance | bring your own data; no proprietary market data is shipped |
BaiQuant is research infrastructure, not investment advice. It is designed for individual researchers who want a local, inspectable workflow for A-share stock selection, strategy experiments, and manual trading review.
- Builds a local A-share research database from your own market data.
- Screens a tradable universe with A-share constraints such as lot size, T+1, limit states, ST names, suspensions, liquidity, and board-specific filters.
- Produces after-close candidate lists for the next trading session.
- Backtests strategy changes and generates charts for review.
- Records manual trades in a local ledger so planned actions can be compared with actual fills.
- Provides a lightweight local Web UI for candidate review, holdings sync, and paper replay.
BaiQuant's public strategy was co-created through an AI-assisted quantitative finance workflow. A human user supplied the trading objective, account size, market observations, screenshots, execution constraints, and risk preferences. OpenAI Codex and GPT-5.5 were then used to turn those requirements into runnable strategy code, tests, backtests, charts, documentation, and a local review UI.
The strategy was not designed as a black-box trading product. It was iterated in a Harness-style evidence loop:
human idea
-> explicit trading rule
-> implementation
-> unit tests and fixture checks
-> historical backtest
-> failure review
-> keep, revise, or reject
That loop matters because most strategy ideas look good in conversation and become fragile once they meet real constraints: delayed after-close data, limit-up/limit-down execution, position sizing, market-regime mismatch, drawdown, and manual order placement. BaiQuant keeps those assumptions visible.
The open-source strategy kept in this repository is steady-20d. It is a
short-cycle A-share stock-selection strategy built around:
- market participation and breadth checks;
- hotspot and industry relative strength;
- multi-factor technical ranking;
- liquidity and tradability filters;
- an approximately 20-trading-day holding horizon;
- next-open execution planning for manual trading;
- individual stops, trailing protection, and portfolio drawdown controls.
The goal is not to predict every day. The goal is to produce a small, reviewable, consistently generated candidate list that a human trader can inspect after the close.
Detailed notes:
The following sample was generated from a local Tushare-backed SQLite database. The strategy was run after the close of 2026-05-25, the first trading day after 2026-05-23. Prices below assume the next session open on 2026-05-26 as the reference execution price, then measure forward movement from daily bars.
| Code | Name | Entry Open | 5D Close Return | 10D Close Return | 20D Close Return | Best High In 20D |
|---|---|---|---|---|---|---|
| 603989.SH | 艾华集团 | 29.82 | +8.08% | +7.14% | +48.96% | +65.29% |
| 000725.SZ | 京东方A | 5.22 | +2.68% | +20.50% | +29.69% | +37.55% |
| 688260.SH | 昀冢科技 | 75.29 | +10.00% | +1.53% | +64.17% | +84.62% |
This is a historical illustration, not an expected return. It may reflect a favorable market window, data-source assumptions, and strategy overfitting. Use walk-forward tests and paper trading before putting real money behind any rule.
Install locally:
python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"Run the test suite:
.venv/bin/python -m pytest -qTry the bundled synthetic fixture data:
.venv/bin/baiquant sample-data --output-dir examples/data
.venv/bin/baiquant run \
--prices examples/data/prices.csv \
--stocks examples/data/stocks.csv \
--fundamentals examples/data/fundamentals.csv \
--events examples/data/events.csv \
--start 2024-01-01 \
--end 2024-01-10BaiQuant does not include proprietary market data. For real A-share research, create your own local database.
Create an environment file:
cp .env.example .envSet your token locally:
export TUSHARE_TOKEN="your-token"Ingest Tushare daily data into SQLite:
.venv/bin/baiquant ingest tushare \
--output data/tushare/baiquant.db \
--start 20230101 \
--end 20260708 \
--adjust qfq \
--write-mode append \
--resumeSee data/README.md for the expected local data layout.
Generate an after-close plan:
.venv/bin/baiquant live20k-2026 \
--db data/tushare/baiquant.db \
--as-of 2026-07-08 \
--preset steady-20d \
--cash 50000 \
--plan-output data/paper/live20k_plan.csvStart the local review desk:
.venv/bin/baiquant desk \
--db data/tushare/baiquant.db \
--host 127.0.0.1 \
--port 8765Record a manual trade in the local ledger:
.venv/bin/baiquant record-trade \
--ledger data/live/trade_log.csv \
--date 2026-07-08 \
--code 600000.SH \
--name ExampleA \
--side buy \
--price 10.00 \
--shares 100The ledger lets you compare planned actions with actual fills during manual review.
src/baiquant/
data/ Data providers and SQLite/CSV adapters
strategy/ Stock-selection and execution planning logic
research/ Factor research and regime validation helpers
desk.py Local Web UI server
live_ledger.py
tests/ Unit and integration coverage
examples/ Small synthetic fixture data
docs/ Strategy and research notes
BaiQuant is smaller and more opinionated than general-purpose quant frameworks. Its focus is the practical after-close workflow of an individual A-share trader. The project is inspired by:
- backtrader, a feature-rich Python framework for backtesting and trading;
- vectorbt, a fast vectorized research and backtesting toolkit;
- Qlib, an AI-oriented quantitative investment platform.
Past performance is not indicative of future results. A-share trading involves market risk, liquidity risk, data-quality risk, execution risk, and model risk. This project is provided for education and research. You are responsible for your own trading decisions.
See DISCLAIMER.md.
MIT. See LICENSE.