Can Large Language Models Execute Parent Orders?
PACE beat standard execution baselines on Shenzhen exchange data without task-specific training.
The paper frames parent-order execution as splitting a large trade into smaller orders while controlling costs. Its proposed LLM-based framework separates long-horizon planning from short-horizon execution and does not require explicit market assumptions. In experiments on Shenzhen Stock Exchange Level-1 data, it outperformed TWAP, Almgren-Chriss, and learning-based baselines, topping the strongest baseline by 0.65 bps. The authors also report that higher model confidence tracked better performance, and that the model tended to trade earlier rather than wait near the deadline. HF Daily Papers' note
The paper frames parent-order execution as splitting a large trade into smaller orders while controlling costs. Its proposed LLM-based framework separates long-horizon planning from short-horizon execution and does not require explicit market assumptions. In experiments on Shenzhen Stock Exchange Level-1 data, it outperformed TWAP, Almgren-Chriss, and learning-based baselines, topping the strongest baseline by 0.65 bps. The authors also report that higher model confidence tracked better performance, and that the model tended to trade earlier rather than wait near the deadline. HF Daily Papers' note
score 4