A Gaussian rule leaves a block-cadence maker short by about 13 percentage points of informed-flow screening when informed traders are scarce. Adadurov, Airapetian, Chtepine, Eloranta, Pletenev, Sebyakin and Valitov put that number on ETH-USDT.
They add a clock to the Copeland-Galai option. The maker quotes against a reference price, then waits for execution before changing the quote. The authors call this interval the quote-exposure horizon h. An informed trader observes the reference move Z_h and trades when |Z_h| exceeds the half-spread s. Each informed order then costs the maker A_h(s), the expected amount by which |Z_h| exceeds s; each uninformed order pays s. Break-even pricing requires s = θ(1−η)/(1−θ)·A_h(s). Here θ is the informed share of potential flow, and η is the fraction of informed orders the maker rejects, its screening rate.
The paper's main object is the screening threshold. Start with the fast venue's break-even spread s0 and price the slow maker's loss at that spread. Matching the venues requires screening away a share η_crit = 1 − A_1(s0)/A_12(s0). An integrable innovation is enough for this result, including innovations with jumps or atoms. Loss-versus-rebalancing, the standard AMM loss measure, scales quadratically with volatility. The paper instead uses a truncated first absolute moment, which remains defined for any integrable innovation without assuming a diffusion. With rare informed flow, the threshold becomes one minus the ratio of mean absolute moves over the two horizons.
For calibration, the authors take the Binance ETH-USDT one-second mid from 1 June to 29 August 2026, covering 7,776,000 seconds. They compare it with 1,725,539 swaps in four Uniswap USDC-WETH pools.
Why does the threshold exceed the Gaussian 71.1%?
At one second the mean absolute move is 0.292 bps; at twelve it is 1.880 bps. Matching the shorter horizon calls for an 84.5% cut, versus 71.1% under a driftless Gaussian. The authors attribute about 3 points of the gap to positive serial dependence: autocorrelation is +0.093, and the realized twelve-second variance by itself implies 74.5%. Distributional shape accounts for the other 10 points. The mid stays still in 73.4% of seconds, while excess kurtosis reaches 154. When it does move, roughly one second in four, the move is about twenty ticks. As a result, κ(1 s) is just 0.43 of its Gaussian value, rising to 0.80 by twelve seconds.
I initially wondered whether the result was counting Binance top-of-book updates rather than executable risk. The authors check this with a microprice reference, which has a zero atom of only 5%, and with a trade-implied mid and the last trade. Across those references, the threshold remains between 84.0% and 85.1%. The calm-second, burst-second pattern persists even when the reference moves almost every second. That makes their case stronger.
The informed share remains the weak input, as the authors acknowledge. Their batch-netting upper bound declines from 0.706 at one second to 0.257 at sixty. At twelve seconds, a run-preserving null accounts for almost all directional imbalance; the excess is −0.007. They therefore sweep θ. Up to θ = 0.257, the threshold ranges from 84.5% to 85.2%. It reaches 87.8% at θ = 0.706, where the Gaussian estimate is 86.4%; the curves cross near θ ≈ 0.78. Such limited movement across a wide sweep is a reasonable stand-in for identification.
The naive spread θσ√h/(1−θ) also runs high against the participation-constrained twelve-second half-spread. At θ = 0.10, it is about 1.7x as large (0.33 versus 0.19 bps); at θ = 0.90, it is about 7x (26.53 versus 3.82 bps).
The pick-off shows up in Uniswap fill prices
In trending seconds, with-trend fills have an effective spread 11.1 bps below against-trend fills after pool fixed effects (t = −27.3). The authors define trending as a trailing sixty-second move above 10 bps. Raw effective spreads are −1.71 bps with the trend and 6.97 bps against it. The sign holds across all four pools. For the 30 bp pools, the gaps are −69.4 and −60.9 bps, about twice the fee. Before with-trend takers arrive, the benchmark has moved 4.85 bps, compared with 2.77 bps ahead of other takers. With pool fixed effects, that difference is 1.77 bps (SE 0.09).
The five-second markout gap is −0.06 bps and insignificant.
Markout-based toxicity dashboards for AMM flow can miss this loss. It has already entered the fill price when the markout clock begins. The authors' fee comparison tells a similar story, though they flag it as not load-bearing. Five-second markout takes up 0.2% to 5.3% of the fee tier. At execution, providers retain a half effective spread of only 1% to 3% of the fee in the v3 5 and 30 bps pools. The corresponding shares are 14% in the v2 pool and 45% in the v3 1 bp pool; the latter accounts for 1,290,417 of the 1,725,539 swaps.
These figures do not establish strategy P&L. The gap groups completed fills using a signal that largely looks backward. In the holdout, twelve-second continuation during trending seconds is 0.130 bps, or 4% of what the trailing drift implies. The in-sample effective-spread regression covers 318,822 trending fills and has an R² of 0.016. It identifies who has already picked off the pool.
Quote age and the threshold
The comparison treats the one-second venue as a subslot design. Slot timing sets the twelve-second pool's horizon, and swaps receive slot-time stamps. As the paper states, nominal block length stands in for h only if it reflects both when the quote binds and when an informed order can reach it. Different age distributions across venues can change the threshold; h_P > h_0 alone does not order their averaged losses. Under uniformly distributed ages, Gaussian loss at zero spread is two-thirds of its endpoint value. If both venues had uniform ages, this factor would cancel from the low-toxicity threshold. The authors caution against treating it as a constant adjustment.
Potential flow stays fixed in the authors' comparison. They cite their own simulations in which subslots increase CEX-DEX arbitrage counts by 535%. A faster venue could draw a different mix of attackers, which remains to be measured. The staleness elasticity is 0.448, close to the diffusion prediction of 0.5. A first-stage F of 3,046 rules out a weak instrument, although missed proposer slots supply the variation and occur in 0.38% of pool-blocks.
We could not reproduce any of this. Our finest price bars span one minute, so the one- and twelve-second horizons both fall within a single bar. That removes the sub-minute calm-and-burst pattern on which the result rests. We also lack pool swap logs, and our data coverage excludes the June to August 2026 window.
A maker budgeting for a twelve-second quote needs a classifier that rejects about six informed orders in seven to match a one-second venue's spread in the low-toxicity limit. Below an informed share of about 0.78, the hurdle exceeds the Gaussian desk estimate. I would next want the threshold calculated from realized quote ages in live subslot data. The authors set the effective horizon, like the informed share, rather than measuring it.