PACE Puts LLMs Into Order Execution, Beats Baselines by 0.65 bps

A new paper on Hugging Face's daily papers list, Can Large Language Models Execute Parent Orders?, takes the LLM-in-finance conversation into a corner that mostly still runs on decades-old math. Not what to trade, but how to trade it, once you already know you have a big order to work.

The setup is called PACE, short for Plan-Ahead Controlled Execution. A Planner reads the parent order and recent market history, forms a long-horizon view, and splits the total quantity across time slots. A short-horizon Executor then decides, minute by minute, how much to push above or below the TWAP baseline. Neither stage is trained on the task. Both are off-the-shelf LLMs called at inference time. The reported experiments use ChatGPT-5.4 and DeepSeek-v4-flash, with Planner allocations weighted against a TWAP anchor by two hyperparameters, lambda and gamma.

On Shenzhen Stock Exchange Level-1 data from April 2026, across 1,680 parent orders, PACE's best variant beats TWAP by 1.02 basis points and the strongest learning-based baseline by 0.65 bps in the aggressive submission setting. Under passive submission the gains are 1.07 and 0.71 bps. Bootstrap resampling puts the aggressive gain over TWAP at p=0.002. For context, on the test set the LLM API cost was about USD 30 to move about USD 35.6 million in traded value; the authors extrapolate that a fund trading USD 100 billion per year would save around USD 10 million annually if it swapped TWAP for PACE. Code is at github.com/zaneopen/PACE.

The behavioural finding is arguably more interesting than the bps. Higher Planner confidence predicted better fills, the opposite of what the human-investor overconfidence literature would suggest, and the Executor front-loaded trades rather than drifting toward the deadline. Sell orders gained more than buys, which the authors attribute to China's short-sale limits leaving more room on that side.

The honest caveat is that this is a backtest on one market, one data vintage, in April 2026 with two specific model checkpoints; live execution introduces latency, queue position, and information leakage that a snapshot replay does not. The reporting also does not tell you how the strategy behaves in a stress event, or what happens to the edge if many desks adopt the same LLM planner and their signals start correlating. If it holds up, though, the direction is what matters: execution algorithms may become another surface where the incumbent modelling tradition gets quietly displaced by prompt-and-call.