Latency is the result a trading desk can use. Passive placement loses 0.011 ticks per contract for every millisecond of round-trip delay, while print-following aggression loses 0.042. Their ordering reverses inside the first 500 microseconds. Passive and aggressive execution cost the same only at zero delay within the tested range, a limitation Maciejewski states directly.
Inside Shadow-PPOV
Shadow-PPOV modifies percentage-of-volume execution by tying placement to observed adds instead of traded volume. Whenever a third party adds on the desired side, the algorithm transmits its own limit order at the same price and venue with fixed probability, 5% in the headline run. It records exactly one item: the followed order's (venue, exchange order id), mapped to its own order id. Size is constrained by 1-lot per-order and 25-lot aggregate limits, the followed order's size and the remaining parent quantity.
When the followed order disappears, the shadow order is cancelled. A delete or modify prompts an immediate cancellation. Following a trade, the algorithm waits through a grace window of end-of-burst events because the sweep that consumed its neighbour may continue into the shadow order.
The economic argument is simple. A resting limit order embodies another participant's decision about where liquidity deserves to be supplied, and Shadow-PPOV adopts that decision without calculating its own placement price. Price-time priority leaves the shadow behind the original order, forcing a sweep to clear the original first. "Model-free is not information-free," the paper says. Its distinction is precise: every message is read, but the algorithm derives nothing from those messages.
The test covers 246 regular sessions from calendar 2025 using ES market-by-order data from CME channel 310. PCAP data is replayed deterministically into a full order-by-order book. Windows begin every ten minutes from 09:29 ET. At each boundary, the algorithm receives 100 contracts to buy and 100 to sell. Completion requires both legs, which leaves slow days with fewer windows of greater length.
Each reported observation averages the two legs against the opening mid, causing that mid to cancel algebraically. This measures a market-making round trip rather than a one-sided parent order. The paper gives a reason for the construction: per-leg intervals are about four times wider than the interval around their own average.
Across 9,592 completed windows, the passive arm costs +0.0955 ± 0.0130 ticks per contract, roughly $1.19 a contract on ES. Immediate crossing at the open would have cost +0.519. The ES book is one tick wide at 96.3% of window openings.
Mark-outs after fills remain flat instead of fading. They measure +0.0825 ± 0.0076 at one second, +0.0743 ± 0.0110 at five and +0.0993 ± 0.0200 at thirty, based on 1.63 million fills. The paper interprets this profile as adverse selection rather than temporary impact. Those mark-outs use the simulator's fill inference, however, so they cannot serve as an independent validation of it.
How much does the zero-latency tie show?
The aggressive comparator uses the same C++ class, triggered by trade prints instead of adds. On 2% of prints, it sends a marketable order at the print price with a 1ms time-to-live. Its cost is +0.0952 ± 0.0135, beside +0.0955 for the passive arm. Measured against immediate execution, the figures are −0.4239 ± 0.0105 and −0.4236 ± 0.0117, respectively.
Both comparisons agree to the fourth decimal, despite one arm filling 1.04% of quantity on arrival and the other filling 51.86%. Participation reaches 4.46% against 4.86%. Those rates were matched by sweeping placement settings rather than targeted directly, because neither arm accepts a requested participation rate.
The tie offers limited support for shadowing itself. Both arms apply the same following mechanism to different message types, leaving order-level shadowing untested against posting at the touch.
VWAP scoring gives passive execution −0.6208 ± 0.0204 and aggressive execution −0.4036 ± 0.0275. The paper explicitly says these values cannot be compared because each leg is measured over its own working interval. Passive legs last 72.7 seconds, versus 62.4. Reading the column as a ranking would be a mistake.
Latency carries the paper
The simulator delays the outbound order, the outbound cancel and the inbound feed together, an appropriate design choice. From zero to 5ms, passive cost climbs from +0.097 to +0.150 ticks per contract, a 55% increase. Aggressive cost rises from +0.087 to +0.296, more than tripling.
At zero delay, aggression is cheaper by 0.010 ticks. By 500 microseconds, it is more expensive by 0.055, with +0.121 against +0.176. The steady deterioration supplies the paper's own evidence that following a specific order conveys information. A policy lacking any timing relationship with order flow would give delay nothing to erode. On the passive side, the inherited information decays at 0.011 ticks per contract per millisecond.
Delay also changes execution style. A passive quote based on a book that has already moved can arrive marketable. Its share filled on arrival therefore increases from 1.1% to 27.3%. Participation moves from 4.51% to 6.87%, while completion time drops from 72 to 48 seconds.
Aggressive execution moves in the opposite direction. Participation declines from 3.18% to 2.02%, and completion time stretches from 92 to 141 seconds. These arms use their own working settings rather than the headline pair: 1% of prints for the aggressive side, with 4.51% participation against 4.46% for passive. The paper clearly limits comparison to rows down a column, rather than across tables.
The 30-second staleness control multiplies passive cost by 3.9, reaching +0.3673 ± 0.0989 from +0.0953.
The configuration remains the same. Participation falls to 1.56%, and each leg records 37.2 child fills rather than 85.5. Legs nevertheless finish in 44 seconds against 72 because stale quotes cross as soon as they arrive. The paper treats thirty seconds as a control, not a plausible latency. Nor does this control establish the value of inherited information relative to placement without any inherited signal.
The missing passive baseline
Maciejewski presents Shadow-PPOV as a reference that predictive placement models should beat. He also writes that "the present study reports no such competing baselines, so the benchmark role is proposed rather than demonstrated." Future work identifies touch posting and the passive legs of TWAP, VWAP and POV as the main comparisons still required.
His defence is that naive alternatives "are ad hoc", whereas Shadow-PPOV "is well specified in a page, has no free parameter that decides where to rest, and can be re-run by anyone with the same data". The placement band weakens the final claim. Tightening that band confines shadowing to adds near the touch, "which fill more reliably", according to the paper. The band therefore acts as a free parameter governing where orders rest, yet its value is absent from the reported runs. No experiment places a touch-posting cost beside +0.0955.
Other constraints are disclosed. The simulator adds algorithmic fills to the day's volume without displacing other orders, leaving every cost pre-reaction. At the 6.9% participation reached with 5ms, the paper calls this assumption strained and says reported participation should be interpreted as an upper bound.
Flow-inherited venue selection is one of five claimed contributions, but the measurements do not exercise it. The paper says the corpus covers a single venue, so this mechanism remains an analytical argument. Parent size stays fixed at 100 contracts.
Size dependence comes from an unfitted Monte Carlo using 4×10^5 trials per size. Its exponential fill gaps ignore correlation across concurrent shadows, as the paper concedes. Relative to the model's own mean, the 10th to 90th participation band spans 0.44 to 9.48 at one child fill and 0.89 to 1.14 at a hundred. The guards receive no ablation. We did not find exchange fees, rebates or message-rate costs in the accounting, a material omission for a rule that places on 5% of adds and cancels each order when its neighbour disappears.
The paper also defends its model-free description. Its guards inspect the inside price and recent trade sizes, which it classifies as risk limits because none forecasts a fill or assigns value to queue position. That defence is mostly fair, apart from the placement band.
Our ES data consists of 1-minute and daily bars. Shadow-PPOV requires per-order adds, cancels and exchange order identifiers, together with queue-position fill inference, so we could not test it.
The paper puts a price on latency for passive orders in a one-tick book, using a full year of message-grain data. That slope gives a desk a direct way to value colocation. Its proposed benchmark status still depends on seeing a touch-posting column beside it.