Can Large Language Models Execute Parent Orders?
Zane Shen, Xinli Xu, Guangyi Zhang, Jialong Chen, Jinsong Zhou, Cong Chen, Guibao Shen, Dongyu Yan, Luozhou Wang, Zhen Yang
cs.CE, cs.CL, q-fin.TR
2026-07-30
PACE splits parent-order execution into long-horizon planning and short-horizon execution and lets an LLM decide at inference, beating the strongest baseline by 0.65 bps on Shenzhen exchange data while trading early and turning confidence into better results.
A core problem in algorithmic trading is parent-order execution. A large order submitted all at once reveals intent and leaves no room to adapt, so it gets worse prices. Parent-order execution splits the large order into smaller slices and decides at each moment how much to trade, aiming to buy lower or sell higher. Two lines of work exist. Static strategies such as TWAP and Almgren-Chriss impose simplifying assumptions about market behavior, like a fixed intraday volume profile, but real markets drift away from those assumptions. Learning-based policies train adaptive rules but need task-specific rewards, states, and actions, and must be retrained for new settings. This paper asks whether LLMs can do parent-order execution, moving LLM use in finance from what to trade to how to execute.
PACE (Plan-Ahead Controlled Execution) is a hierarchical framework with no market assumptions and no task-specific training; decisions are generated at inference. It splits execution into two layers. Inputs are the parent order (stock, direction, window, total quantity), market history (price and volume over the past L minutes, using the mid-price of the best bid and ask as the price proxy), and the TWAP curve (quantity spread evenly across decision points, the simplest baseline).
The Planner handles long-horizon planning. It produces a textual trend assessment over the execution window, a set of allocation scores, and an overall confidence c. The scores blend with the uniform TWAP allocation, where lambda controls the LLM share: larger means more model, smaller means closer to TWAP. This splits the total quantity into N sub-plans. The Executor handles short-horizon execution: given a sub-plan, it combines market history, the Planner long-horizon trend, and the TWAP baseline to tweak quantity at each decision point (controlled by gamma), trying to trade more when the local price is favorable. Backtests run on Shenzhen Stock Exchange Level-1 data. Models tested are GPT-5.4 and DeepSeek V4 Flash (DS-v4-f).
Across 1,680 parent orders (total traded value about USD 35.6 million), PACE beats every static and learning baseline, and all strategies complete 100% of execution.
| Strategy | Type | Aggressive wbp | Gain over TWAP |
| TWAP | static | -3.28 | — |
| Almgren-Chriss | static | -2.93 | +0.35 |
| XGBoost | learning | -3.10 | +0.18 |
| LSTM | learning | -2.91 | +0.37 |
| GPT-5.4 | LLM | -2.76 | +0.52 |
| DS-v4-f | LLM | -2.26 | +1.02 |
The strongest variant, DS-v4-f, beats TWAP by 1.02 bps under aggressive submission (95% CI 0.15 to 2.12, p=0.002) and the strongest baseline by 0.65 bps; under passive submission the gains are 1.07 and 0.71 bps (p=0.014). At 1 bps per USD 100 billion traded per year, roughly USD 10 million, the authors estimate 0.65 bps as about USD 6.5 million in annual savings. The total LLM API cost for the whole experiment is about USD 30.
Gains hold across groups: sell orders gain more (China short-sale limits let negative information enter prices slowly, leaving execution room), and both low- and high-volatility groups win, with PACE more noise-robust than learning strategies. Ablation shows removing the Planner or the Executor hurts, and the Planner contributes more. Removing side-specific guidance or the market glossary also regresses, yet the model retains execution ability even without them.
Several regressions show the LLM behaves unlike a human. Higher confidence predicts better performance: the Planner confidence c has a significantly positive coefficient on bp (p<0.01), opposite to human overconfidence hurting returns (Odean 1999). The model trades early: the Executor places quantity when time pressure is low, and TP has a significantly negative coefficient, opposite to human procrastination near deadlines (Steel and Konig 2006). DS-v4-f decisions are not explained by simple trend-following or mean-reversion (recent-return LR is insignificant), while GPT-5.4 partly follows recent trends.
It is the first systematic test of LLMs on parent-order execution, a real problem with a small but economically meaningful effect and trivial inference cost. The behavior findings matter more: the LLM is more disciplined than humans at execution, trades early when it should, and converts confidence into better results, pointing to a role of complementing human traders rather than replacing them.
This is a backtest, not live trading, which the authors flag as future work. The effect size of 0.65 bps is small, and whether it survives real frictions, impact costs, and regime shifts is a question for live trading. The market and asset are single (Shenzhen A-shares), and inputs are only price-volume and TWAP, with no news, microstructure, or cross-asset signals. Scoring uses simulated backtest fills, which differ from live fills.