PACE beats TWAP by 1.02 bp with DeepSeek-v4-flash and all-aggressive submission, within the simulator used for the paper. I distrust the simulator. Once Shen and coauthors limit fills to displayed best-quote volume, the advantage drops to 0.67 bp. This depth-limited run is the paper's only test that restricts fills to available quote size, and it removes a third of the TWAP-relative gain. Almgren-Chriss, XGBoost and LSTM were not run again under the same rule. The abstract's +0.65 bps over the strongest baseline therefore remains untested against depth.
PACE (Plan-Ahead Controlled Execution) receives a parent order with its stock, side, quantity and window. It divides the choice between two components. The long-horizon LLM Planner takes one lookback window of minute mid-prices and volumes, scores every fixed 5-minute slot from -1 to +1 on a five-point scale, and assigns confidence c between 0 and 1. Those scores alter the flat TWAP schedule. Each slot receives (1 - lambdac)/N of TWAP plus lambdac times a softmax of the scores. With lambda = 0.3, confidence determines the distance from a flat schedule. Inside each slot, the short-horizon LLM Executor issues a per-minute score z on that same five-point scale. It trades (1 + gamma*z) times the TWAP clip, with gamma = 0.5. Each order requires an LLM call every minute.
PACE never selects the price. The paper instead fixes pricing in one of two settings. Under all-aggressive submission, buys cross at Ask1 and sells cross at Bid1. Under all-passive submission, orders rest at the opposite touch, with any residual swept aggressively in the final minute. PACE is therefore a quantity schedule placed over TWAP, with gains expected from shifting volume toward better minutes. The performance measure is value-weighted price performance in bps relative to the average mid during the window.
The sample uses Shenzhen Stock Exchange Level-1 snapshots from all trading days of April 2026. Ten parent orders are generated randomly each day and run eight times, producing 1,680 runs from roughly 210 distinct orders. Order quantities range from 100 to 10,000 shares, with a mean of 5,000.48. Windows run from 10 to 60 minutes and average 35.29, while 51% are buys.
In the aggressive setting, DeepSeek-v4-flash records -2.26 bps against -3.28 for TWAP. The differences are +1.02 bps over TWAP and +0.65 over the strongest baseline, LSTM at -2.91. Passive results are -3.92 for DeepSeek-v4-flash and -4.99 for TWAP, giving +1.07 over TWAP and +0.71 over XGBoost at -4.63. ChatGPT-5.4 delivers about half the improvement, +0.52 and +0.50. Every strategy completes 100%.
We could not test any of these results. The paper trades Shenzhen equities, and its matcher requires Ask1 and Bid1. Our data consists of minute OHLCV for US names, without equity quotes, so we cannot reproduce fill logic tied to the touch. Substituting the LLMs available to us for the two models used in the paper would also make the model-specific results incomparable.
How much does the depth cap cost?
In the main matcher, a buy fills whenever its price reaches or exceeds Ask1. Algorithm 1 fills the entire order, without a size cap or queue. The matching rule produces the main table's 100% completion rate. Market impact, the problem parent-order execution is designed to manage, is absent by construction.
The authors also apply a stricter rule. With depth-limited matching, aggressive fills cannot exceed Ask1/Bid1 volume, and the unmatched remainder is cancelled. PACE then returns -2.41 while TWAP returns -3.08. The authors say the result "shows that PACE still outperforms TWAP under this more conservative matching mechanism." Their statement is accurate. PACE gets worse and TWAP gets better, leaving 0.67 bps of the gain, 34% below 1.02. The direction fits an edge partly generated by PACE concentrating more size in individual minutes than the touch can take.
The capped test reports no completion rate, and the authors rerun no baseline other than TWAP. Its cap covers displayed best-quote size alone, leaving a multi-level book and adverse selection on passive orders outside the test.
Order size reaches the same issue from another direction. Across the 1,680 parent orders in the main experiment, traded value totals about USD 35.6 million, or roughly USD 21k per order. The paper's scaling note values 1 bps on that notional at USD 3,560, compared with about USD 30 in API cost. It then projects USD 10 million annually for a fund trading USD 100 billion. A USD 21k parent order has no impact problem. A fund trading USD 100 billion has little else.
One month of front-loading
The cross-section breaks my confidence in the result. The Executor consistently front-loads: every specification gives the time-pressure coefficient a negative sign, with t reaching -11.42. In Figure 3, aggressive DS-v4-f gains over TWAP are +1.43 bps for sells and +0.35 for buys. Short windows gain +1.45, compared with +0.62 for the 30 to 60 minute group. Even the Almgren-Chriss baseline gains +0.35 bps. Here it has been reduced to a one-parameter sinh front-loading curve, with kappa = 0.5 selected by grid search.
Early trading helped during April 2026. In those aggressive DS-v4-f bars, the benefit for sells was about four times the benefit for buys.
Shen and coauthors attribute this asymmetry to short-sale limits delaying the incorporation of negative information in China, citing Chang, Luo and Ren. Short-sale limits offer a real mechanism. The same pattern would arise if prices simply drifted down after 10:30 during the single month of April 2026. One month, one fixed start at 10:30:00, and windows that never include the open or close cannot distinguish those explanations. A second month with drift in the opposite direction would settle the question. That result would change my mind.
PACE's lambda, gamma and tau underwent sensitivity scans on the April 2026 evaluation data. The baselines received equivalent treatment: AC's kappa and the ML adjustment eta were each selected as the winner from a five-point grid. Every tuning decision uses the evaluation month. It is the same hindsight-parameter problem discussed in an earlier note on Wasserstein-ball allocation.
The bootstrap supplies what evidence it can. Across 5,000 parent-order resamples, the aggressive estimate is +1.02 bps with a 95% interval of [0.15, 2.12] and p = 0.002. The passive estimate is +1.07 bps with [0.05, 2.24] and p = 0.014. Its lower bound is 0.05 bps. GPT-5.4 intervals begin at 0.00, and the paper acknowledges that it "remains close to the significance boundary."
The behavioural evidence is worth keeping
Planner confidence has a positive relationship with realized bp in all six regressions. Coefficients range from 4.78 to 7.34, with t between 2.04 and 2.58 and N = 1,680. Standard errors are clustered by parent order, the appropriate correction given eight repeats for each order. Adjusted R-squared runs from 0.020 to 0.041. The direction is useful; the fit offers nothing to size on.
The Executor regressions include the side-adjusted recent log return. GPT-5.4 has a strongly positive coefficient of 10.49, with t = 4.86. DeepSeek produces 1.72, with t = 0.53, an insignificant result. The authors interpret that insignificant coefficient as evidence that simple trend-following or mean-reversion cannot account for DS-v4-f's choices. They combine it with the ablation, in which removing the Planner still beats TWAP by +0.40 bps, and argue that these non-heuristic decisions generate genuine performance gains. Adjusted R-squared for the Executor regressions ranges from 0.013 to 0.097. Claiming a better signal from one insignificant regressor in a regression that explains almost nothing rests on weak ground.
Every strategy in Table 1, including PACE, loses against the mid benchmark. DeepSeek-v4-flash loses 2.26 bps under aggressive submission. It recovers 1.02 of the 3.28 bps that TWAP pays to cross. The -2.26 result comes from a matcher that disregards available size. In the sole test with a size cap, 0.67 bps remains, and the authors are right that PACE continues to beat TWAP there.
The two objections with force come from the authors' own work. Their depth cap supplies the stricter matching test, and their discussion of GPT-5.4 identifies model instability. The conclusion also concedes that microstructure inputs are absent and that the method has never run live. The abstract's unqualified "outperforms" deserves pressure. Until another month with intraday drift in the opposite direction produces the same result, this remains a scheduling experiment.