Can Large Language Models Execute Parent Orders?

Battle Count: 72

Directly addresses a core problem in algorithmic trading (parent-order execution). Demonstrates that LLMs can outperform traditional (TWAP, Almgren-Chriss) and ML-based (XGBoost, LSTM) execution strategies. The 0.65-1.07 bps improvement translates to USD 6.5-10 million annual savings for a USD 100 billion fund. Behavioral analysis provides novel insights into LLM decision-making patterns relevant to execution. The framework requires no task-specific training, making it adaptable to new market settings.

The framework is conceptually straightforward with two LLM calls (Planner and Executor) per parent order. Implementation requires: (1) structured prompt engineering with JSON output parsing, (2) a backtesting environment with order matching logic, (3) hyperparameter tuning (lambda, gamma, tau, delta), (4) LLM API integration with error handling for stochastic outputs. The main complexity lies in prompt design and ensuring reliable JSON parsing from LLM outputs. No model training is required, reducing implementation burden compared to RL-based approaches.

Code is publicly available on GitHub (https://github.com/zaneopen/PACE). The paper provides detailed hyperparameters (lambda=0.3, gamma=0.5, tau=5min, delta=1min), prompt templates, matching algorithm (Algorithm 1), and parent order generation procedure. However, it relies on proprietary LLM APIs (ChatGPT-5.4, DeepSeek-v4-flash) which may change over time. Data is from Shenzhen Stock Exchange Level-1 snapshots for April 2026, which may not be freely accessible.