If you post passive size in a Shenzhen stock, worry about orders that hit you at exactly your price. This is the most useful finding in Guo, Li, Qi and Sun, and the part of their paper that stands up best.
The mechanism is straightforward. Retail apps display snapshots of the Shenzhen book every three to six seconds. Traders who reconstruct it from Level-2 order messages can see it at millisecond resolution. During the first hour, best quotes change 1.327 to 7.053 times per second across size groups. Of those updates, 51.52% to 64.90% come within 100ms of the preceding update. An incoming limit order priced exactly at the best opposite quote therefore reveals something about its sender: the trader was watching a current book. The authors name this an active exact match. Their auxiliary measure, a passive exact match, identifies orders priced at the best same-side quote.
The data cover July 2023 to October 2025, excluding the bottom 30% of stocks by size. The authors reconstruct 47,418,470,233 orders in continuous trading, then take a 0.1% systematic sample containing 42,506,216 limit orders across 2,760 stocks and 566 days. The main result concerns prediction. After controlling for direction, aggressiveness and order-size traits, active exact matches precede a 2.157 bp higher direction-adjusted mid-quote return over ten seconds (t=39.20). The paper then turns to market quality, using congestion at the open as a latency shock.
Why should an exact hit reveal speed?
The behavioral evidence is stronger than I expected. Active matches are 19.2 points more likely to arrive within 100ms of a quote change in the full-control regression. For canceled orders in the largest size quintile, their median life is 3.01 seconds; everything else lasts 36.04 seconds. Returns provide the clearest distinction from ordinary aggressiveness. Orders priced more than a tick through the opposite quote have a -2.828 bp coefficient, and those exactly one tick through have -0.435 bp. Exact matches carry +2.157 bp, with all coefficients measured against non-matching orders. Paying up suggests impatience. Hitting the price to the tick suggests someone is following the book.
The tick-quote test reinforces that reading. Its separate specification uses integer-second return endpoints and its own snapshot sample, so these estimates are not directly comparable with the 2.157 bp baseline. Matching the stale three-second snapshot has -1.248 bp of incremental predictability. Matching the millisecond book while missing the three-second snapshot carries 3.797 bp. Matching both contributes a further -1.028 bp.
Scale is the weakness. Active matches account for 27.63% of selected orders. Assigning over a quarter of all order flow to latency-sensitive firms stretches belief in a retail-dominated market. Our inference is that some of this flow likely consists of retail orders priced at the ask. Linking exact-match activity to trader-level identifiers remains future research in the authors' account. Behavior, returns and the response to latency validate the proxy; trader-level identifiers have never been used to validate it.
Fast matches, bounded by displayed depth
The speed split locates much of the information. Once a 100ms response dummy and its interaction enter the regression, the plain active-match coefficient falls from 2.157 to 0.427 bp. The interaction contributes 2.174 bp (t=35.47). Speed by itself earns 2.182 bp (t=47.00) across orders, whether they match or not. Late active matches retain some predictive content. Most of it belongs to matches that arrive fast.
Size relative to displayed depth gives the other useful split. Matches smaller than opposite depth predict 2.711 bp; exact-size matches predict 1.818 bp. Above that depth, the prediction turns to -4.408 bp (t=-68.39). The authors interpret this as inventory discipline, since a high-frequency trader is less likely to oversize.
A mechanical explanation is also available. The residual from an oversized order rests at the new same-side best. Because returns begin at the first valid post-submission midpoint, that starting price may already reflect the residual. A subsequent reversal could follow from that. I would want the authors to repeat this split using the midpoint before submission as the starting price.
The persistence surprised me. Across a common sample of 39,509,310 orders, the coefficient rises from 1.671 bp at 100ms to 3.747 bp at 600s.
A signal native to milliseconds predicts ten-minute drift.
A brief advantage at the open
The latency design is clever. Shenzhen allocates stocks to ten channels, and opening congestion can hold up order messages for more than ten seconds. Channels clear at different times. On the example day, October 20, 2025, delays in the fast and slow groups fall below three seconds at 38s and 47s after the open. The comparison uses the interval between those crossings.
Active share increases 0.505 pp (t=3.03), with a clean pre-trend (p=0.228). Active count increases 0.447 orders per stock-second, although its pre-trend rejects at p=0.028. The authors accordingly say they "treat the count result as supporting evidence." One-second adverse selection increases 0.648 bp as realized spread declines 1.345 bp. These belong together as a single finding: realized spread equals effective spread minus twice adverse selection. Effective spread barely changes (-0.049 bp, t=-0.53).
Wide-spread incidence increases 2.964 pp. Yet the share of time spent wide on the authors' fixed 100ms grid increases just 0.520 pp (p=0.0527). Quotes wider than one tick occur more frequently. Our reading is that they do not persist.
The exercise covers 24 usable days out of 36, confined to seconds after 09:30. Assigning these effects specifically to HFT rather than faster data generally depends on an exclusion restriction. The authors state it. They do not test it.
The 10.95% a year leaves out the spread these orders cross
Averaging active exact-match orders' ten-second returns by stock-day gives the authors 4.125 bp a day, or 10.95% a year. Their qualification is explicit: "these figures illustrate the magnitude of the predictive signal rather than realizable trading profits." They identify the missing costs and constraints themselves: spreads, commissions, transaction tax, short-sale limits and T+1.
Spreads are the sharpest problem here. Active matches cross the spread by construction, whereas the 2.157 bp measures a mid-to-mid return. Anyone following the flag pays the half-spread to enter and again to exit. Mean ten-second returns for active matches range from 3.1379 bp in the largest size quintile to 4.9228 bp in the smallest. That drift must pay for both legs before taxes.
We could not test any of this. Constructing the flag requires Shenzhen order-by-order messages and a reconstructed book at millisecond resolution. We hold US one-minute bars without quotes. Those bars cannot reveal an order's price relative to the book at arrival.
The practical application is defensive. Passive liquidity providers in Shenzhen should regard fast orders hitting their exact price as toxic. Quote-matching orders smaller than displayed depth carry +2.711 bp at ten seconds; the pooled active-match coefficient reaches 3.747 bp at 600s. Converting that flag into P&L requires a net-of-spread test, which the paper explicitly says it does not attempt.