TradeGrad

This is the implementation of our paper Trading Strategy Optimization via Textual Gradient.

TradeGrad uses an LLM to iteratively rewrite a trading strategy's source code, guided by textual feedback derived from its own backtest performance. The optimizer, TradeTGD, samples a reference set from the score-ranked strategies tried so far, writes a textual gradient against that set plus a running memory summary, picks the next base program uniformly from the references, and revises it as either a small local edit or a full rewrite. Training is scored by the worst-30% CVaR over yearly backtest windows, so the search can't win by blowing up in a bad year; a trade-frequency gate rules out the degenerate "stop trading to flatten drawdown" solution.

Layout

src/               optimizer (TradeTGD), loss/reward, backtest evaluators, task registry, run loop
seed_strategies/   initial strategy each task starts from
script/            train.sh / eval.sh -- entry points, see below

Setup

pip install -r requirements.txt

Two things live outside this repo and need to be pointed at on your machine:

• OPENROUTER_API_KEY -- or any OpenAI-compatible key (see --base-url/--model on run_one.py)
• AKQUANT_DATA_ROOT -- per-symbol OHLCV csvs and index constituent lists (default: a data/ directory next to this repo). Expected layout:

  <AKQUANT_DATA_ROOT>/CN/<symbol>.csv         <AKQUANT_DATA_ROOT>/US/<symbol>.csv
  <AKQUANT_DATA_ROOT>/CN_index/sh.000300.csv  <AKQUANT_DATA_ROOT>/US_index/SP500.csv
  <AKQUANT_DATA_ROOT>/hs300_list/*_list_*.csv <AKQUANT_DATA_ROOT>/SP500_list/*_list_*.csv
  

Running

script/train.sh and script/eval.sh hold their settings as plain variables at the top of the file (YOUR_DATA_ROOT / YOUR_API_KEY placeholders included) -- edit them, then:

bash script/train.sh      # train a task, then score it on the held-out 2024-2025 window
bash script/eval.sh       # re-score an existing run without retraining

Both refuse to run while a placeholder is unedited, and both launch detached (nohup, backgrounded) so the run survives the terminal disconnecting. script/my_train.sh / script/my_eval.sh are gitignored personal copies with your machine's settings filled in, for everyday use.

Each run writes to result/<MARKET>/<mode>/checkpoint/<step>/, checkpoint_test/<step>/ (the same checkpoints scored on the test window), and best/ (the highest-scoring checkpoint once training finishes). Neither result/ nor log/ is tracked by git.