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

Source-linked AI summary

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

arXiv:2607.28410v1Jul 30, 2026cs.CEcs.CLq-fin.TR

TL;DR

Parent-order execution requires splitting large orders to reduce execution costs, but existing methods depend on market assumptions or task-specific training. This paper introduces PACE, an assumption-free, training-free hierarchical LLM framework that outperforms execution baselines by up to 0.65 bps and exhibits distinct trading behaviors.

Problem
Parent-order execution must determine how much to trade over time to reduce costs, while existing methods rely on market assumptions or task-specific training and LLM research has focused on what to trade.
Method
PACE hierarchically separates parent-order execution into long-horizon planning and short-horizon execution without pre-specified market assumptions or task-specific policy training.
Results
PACE consistently outperforms all baselines across both order-submission settings, exceeding the strongest baseline by 0.65 bps.
Takeaways & Limitations
Behavioral analysis suggests LLMs may complement human traders by making execution decisions with confidence linked to better performance and by trading earlier rather than deferring execution.

Abstract

from arXiv · show

Parent-order execution is a core problem in algorithmic trading, where the goal is to split a large order into smaller orders while reducing execution costs. Existing approaches either rely on pre-specified market assumptions that may not hold in practice, or require task-specific training that limits adaptability to new settings. To overcome these limitations, we present the first systematic study of large language models (LLMs) for parent-order execution. This extends the use of LLMs in finance from what to trade to how to execute. We propose PACE (Plan-Ahead Controlled Execution), a hierarchical framework that decomposes parent-order execution into long-horizon planning and short-horizon execution, requiring neither explicit market assumptions nor task-specific training. Experiments on Shenzhen Stock Exchange Level-1 data show that PACE outperforms TWAP, Almgren-Chriss, and learning-based baselines, exceeding the strongest baseline by 0.65 bps. Behavioral analysis reveals that LLMs make execution decisions differently from human investors: higher model confidence predicts better performance rather than worse returns, and the model trades earlier rather than procrastinating toward the deadline. These findings suggest that LLMs can complement human traders in execution decisions.

1 Introduction

Parent-order execution reduces the costs and risks of splitting large orders without relying on fragile market assumptions or task-specific training. The paper introduces PACE, which combines long-horizon planning with short-horizon execution, outperforming established baselines and exhibiting decision patterns that may complement human traders.

Motivation: Parent-order execution splits large orders into smaller trades to reduce information leakage, adapt to market changes, and limit execution costs.Submitting a large order at once can reveal trader intent, worsen traded prices, and leave no room for adjustment.
Limitations of Existing Strategies: Traditional strategies rely on simplifying market assumptions, while learning-based strategies require task-specific rewards, states, and actions that may need redesign when conditions change.The paper identifies the need for strategies that avoid both pre-specified assumptions and task-specific training.
PACE: PACE decomposes parent-order execution into a long-horizon Planner and a short-horizon Executor, requiring neither pre-specified market assumptions nor task-specific training.The Planner generates long-horizon plans, while the Executor adjusts trading quantities in response to short-horizon market changes.
Empirical Results: 0.65 basis points (bps) is PACE’s improvement over the strongest baseline on Shenzhen Stock Exchange Level-1 data, corresponding to USD 6.5 million in annual savings for a fund trading USD 100 billion yearly.The evaluation compares PACE with TWAP, Almgren-Chriss, XGBoost, and LSTM on identical parent orders.
Behavioral Analysis: Behaviorally, higher Planner confidence predicts better performance, and the Executor trades earlier rather than waiting until deadlines, unlike the cited human patterns.These differences suggest that LLMs may complement human traders in complex execution decisions.

2 Related Work

Prior parent-order execution research follows static and learning-based approaches, while PACE introduces LLMs to the task for the first time. More broadly, financial LLM research spans financial language models and resources alongside applications of LLMs and foundation models to quantitative tasks.

Parent-order execution: Existing parent-order execution work mainly follows static and learning-based lines.Static methods formulate execution through market-impact modeling or stochastic control, with Almgren–Chriss cited as a representative mean-variance framework.
Parent-order execution: PACE is the first approach to introduce LLMs to parent-order execution, combining pretrained knowledge with current market observations for adaptive inference-time decisions.It does so without specifying a price model or training a task-specific policy.
LLMs in finance: Existing financial LLM research covers financial language models and resources, plus applications of LLMs and foundation models to quantitative tasks.Examples of financial language models and resources include FinBERT, FLANG, BloombergGPT, FinGPT, and PIXIU.

3 Method

PACE is a hierarchical parent-order execution framework that combines long-horizon LLM planning with short-horizon execution, using TWAP as a controllable reference. Its backtesting environment evaluates submitted orders through market matching and execution metrics.

Inputs: Parent-order execution splits a large order into smaller trades to reduce liquidity consumption, information leakage, and execution costs.A parent order specifies the stock ID, trading direction, execution window, and total execution quantity.
Inputs: PACE receives the parent order, market history, and TWAP curve as inputs.The parent order is O = (s, d, ts, te, Q), with execution window [ts, te] and duration T = te − ts.
PACE framework: PACE separates long-horizon planning from short-horizon execution to address the limited adaptability of static strategies and the limited prior knowledge of learning-based strategies.The Planner forms a textual trend assessment over the execution window and decomposes the parent order into shorter execution sub-plans.
PACE framework: The Planner assigns each time slot a quantity preference score an ∈ [−1, 1] and an overall confidence score c ∈ [0, 1], then mixes LLM allocation with TWAP through λ ∈ [0, 1].A larger λ increases the weight on LLM allocation, while a smaller λ keeps allocation closer to TWAP.
PACE framework: The Executor adjusts the TWAP baseline using market history, the Planner’s trend assessment, and zt ∈ [−1, 1], with γ ∈ [0, 1] controlling deviation from TWAP.zt = 0 follows TWAP; positive and negative values increase and decrease the current order quantity, respectively.
Backtesting environment: The backtesting environment uses a Matcher to determine fills and an Evaluator to compute strategy metrics from traded orders.Aggressive orders are submitted at prices allowing immediate trading, whereas passive orders seek better prices but may remain unfilled.

4 Experiments

Experiments on Shenzhen Stock Exchange Level-1 data show that PACE consistently outperforms benchmark strategies, with statistically significant gains and benefits from both planning and execution modules. Behavioral analyses further indicate stable, confidence-sensitive, and non-procrastinating execution decisions.

Main Results: PACE consistently outperforms all baselines, improving wbp over TWAP by 1.02 bps aggressively and 1.07 bps passively.The corresponding improvements over the strongest baseline are 0.65 bps and 0.71 bps.
Statistical Significance: 1.02 bps and 1.07 bps gains over TWAP are statistically significant under aggressive and passive submission, respectively.The 95% confidence intervals are [0.15, 2.12] and [0.05, 2.24], with p = 0.002 and p = 0.014.
Heterogeneity: PACE’s gains are larger for sell orders and smaller for longer execution windows.The paper attributes larger sell-side gains to China’s short-sale limits and smaller long-horizon gains to reduced price predictability.
Module Ablation: Removing either PACE module hurts performance: w/o P removes planning, w/o E removes execution, and w/o P+E equals TWAP.The Planner contributes more in the all-aggressive setting, while the Executor adds clear value beyond uniform execution.
Case Study: The Planner front-loads sell quantity before a realized price decline, while the Executor increases a sell allocation from the TWAP baseline of 200 shares to 300 shares.The case study uses only pre-execution history for planning and information available at each decision time for execution.
Behavioral Analysis: Higher confidence produces more concentrated allocations, and negative time-pressure coefficients show that the Executor trades more earlier rather than procrastinating near deadlines.The confidence relationship has positive slopes across all three model settings, while the earlier-trading pattern differs from typical human deadline behavior.

5 Conclusion

This paper presents the first study of LLMs for parent-order execution and proposes PACE, which separates long-horizon planning from short-horizon execution without pre-specified market assumptions or task-specific training. Experiments show consistent improvements over all baselines, while behavioral analysis finds that LLM decisions differ from those of human investors.

The paper presents the first study of LLMs for parent-order execution.
PACE separates long-horizon planning from short-horizon execution without pre-specified market assumptions or task-specific training.
Experiments demonstrate consistent improvements over all baselines under both order-submission settings.
Behavioral analysis reveals that LLM decisions differ markedly from those of human investors, with higher model confidence predicting better performance.

6 Appendix

The appendix describes the parent-order evaluation setup, the Almgren–Chriss baseline implementation, and robustness checks for PACE. These checks use bootstrap resampling and stricter matching conditions to assess whether PACE’s gains persist.

Evaluation setup: Each trading day uses 10 randomly generated parent orders, with a fixed 10:30:00 start and execution windows sampled from 10 to 60 minutes.End times therefore range from 10:40:00 to 11:30:00, while order side is randomly sampled as BUY or SELL.
Robustness analysis: 5,000 bootstrap resamples test PACE’s statistical robustness, with DS-v4-f showing positive 95% confidence intervals under both order-submission settings.GPT-5.4 remains close to the significance boundary, while the results support the statistical robustness of the strongest PACE variant.
Robustness analysis: PACE is also evaluated under a stricter depth-limited matching rule to test whether its effectiveness remains unchanged under more restrictive matching.The supplied passage introduces this test but does not state its outcome.
Baselines: The Almgren–Chriss baseline computes target remaining inventory from parent-order quantity Q and execution-window length T.The implementation defines target remaining inventory at time t as its first step.

AC Strategy

The AC strategy controls execution front-loading through κ: κ = 0 produces an approximately uniform TWAP curve, while larger κ values execute more quantity earlier. A grid search selected κ = 0.5 as the best-performing parameter.

AC Strategy: κ controls front-loading: κ = 0 yields an approximately uniform TWAP curve, while increasing κ executes more quantity earlier.The strategy determines target cumulative executed quantity at each minute from the target remaining inventory.
AC Strategy: κ = 0.5 was selected as the best-performing parameter after searching κ ∈ {0.1, 0.3, 0.5, 0.8, 1.0}.

Learning-based Strategies

Learning-based strategies train on market-wide minute-level data to predict next-minute price direction and adjust TWAP execution quantities accordingly. The framework uses a tunable adjustment magnitude and evaluates Planner stability through repeated quantity-preference queries.

Data and training: Models train on market-wide minute-level data from January 1, 2026 to March 31, 2026 and evaluate on April 2026 data.The training period precedes the evaluation period.
Data and training: Inputs include log return, log-return volatility, order imbalance, trade imbalance, and spread volatility, with targets labeling next-minute returns as up, down, or flat.The prediction target is a three-class direction label.
Execution rule: The strategies use TWAP as the default execution curve and increase current TWAP quantity when predicted price movement favors the parent order’s direction.For buys, quantity increases on upward predictions; for sells, it increases on downward predictions.
Execution rule: The adjustment magnitude is controlled by η, searched over η ∈{0.1, 0.3, 0.5, 0.8, 1.0}, with the best result reported in the main experiments.The signal indicator is y_t = 1 for favorable current signals and y_t = 0 otherwise.
Planner stability: Repeated-query testing examines whether the Planner produces stable long-horizon quantity-preference scores for the same parent order.For parent order i and run r, outputs are slot scores a_i,r converted into an allocation distribution using a fixed softmax transformation.

Planner Stability

Planner stability is evaluated across R = 8 repeated runs by comparing each run’s score distribution with the parent-order average. Lower KL dispersion indicates more similar long-horizon score distributions across runs, while score concentration is measured to examine its association with Planner confidence.

For each parent order, the analysis computes an average score distribution across R = 8 repeated runs.
KL dispersion measures the divergence between each run’s score distribution and the parent-order average distribution.
Lower KL dispersion means the Planner produces more similar long-horizon score distributions for the same parent order across repeated runs.The model-level statistic is the average KLDispersion_i across parent orders.
Score concentration is measured with the HHI to examine whether higher Planner confidence corresponds to more concentrated long-horizon allocation preferences.

Confidence and Concentration

This section examines whether Planner confidence predicts execution performance and how confidence relates to the concentration of score-implied allocations. It uses demeaning to compare repeated runs relative to each parent order’s average.

Confidence and Concentration: A larger HHI indicates that the Planner’s long-horizon preference is concentrated on fewer slots.HHI measures concentration in the Planner’s long-horizon slot preferences.
Confidence and Concentration: Demeaning confidence and HHI by parent-order means compares whether relatively more confident runs also have more concentrated score-implied allocation distributions.This controls for differences across parent orders when comparing repeated runs.
Confidence and Performance: Planner confidence score c is used as the main explanatory variable, with bp as the dependent variable for execution performance.The analysis tests whether confidence predicts execution performance.

Confidence and Performance · Executor Regression

The paper examines how confidence and market conditions relate to the Executor’s quantity adjustments using controlled OLS specifications. It also defines time pressure and recent side-adjusted returns for the Executor regression and describes the Planner and Executor prompt inputs.

Confidence and Performance: The Confidence and Performance analysis estimates OLS specifications with and without controls.The controls include parent-order side, log parent-order quantity, execution-window length, and stock volatility.
Confidence and Performance: The analysis uses the LLM-generated quantity adjustment score z, where larger z indicates greater allocated quantity.The regression examines responses to time pressure and recent price movements.
Executor Regression: Executor Regression defines time pressure TPt so larger values indicate that the current time is closer to the sub-plan deadline.t denotes remaining sub-plan time at time t, while τ denotes total sub-plan duration.
Executor Regression: Executor Regression defines LR as the side-adjusted log return over the previous τ minutes using the parent-order direction.For BUY orders it uses log Pt − log Pt−τ, while SELL orders reverse the sign; Pt is the mid-price.
Executor Regression: The Executor Regression reports OLS specifications with and without controls in Tab.5.The controls add log sub-plan quantity to parent-order side, log parent-order quantity, execution-window length, and stock volatility.
6.6 Prompts: The paper presents the prompt details for the Planner and Executor in Figure 8 and Figure 9.Figure 8 shows the condensed Planner prompt template, while Figure 9 shows the condensed Executor prompt template.
6.7 AI Assistance Statement: The Planner scores fixed future slots using minute-level historical mid-price and volume together with parent-order direction, quantity, duration, and fixed 5-minute slots.Runtime market history, parent-order fields, and slot definitions are filled dynamically for each parent order.
6.7 AI Assistance Statement: The Executor adjusts the current minute’s quantity using recent market information, Planner intent, the TWAP baseline quantity, and execution progress.Inputs include current time, remaining minutes, executed and pending quantities, and recent minute-level mid-price and traded volume.