Ethresear.ch: a 12-second AMM must screen 84.5% of informed flow to match a 1-second venue
Serguey Asael Shinder3 min readA research post published on ethresear.ch on 15 October 2026 by the author nuconstruct asks a question that matters for anyone providing liquidity on Ethereum: how much is a second of reduced quote staleness worth, compared with turning away informed traders? The post builds a break-even model for a market maker and takes it to on-chain data.
The set-up. A maker that posts a firm two-sided quote, the author writes, has effectively written an option: traders who know where the price is heading exercise it, and the maker recovers those losses through the spread it charges everyone else. Automated market makers make this stark, because a constant-function pool only refreshes its price when someone trades and, on a chain that settles in blocks, can stay stale for the whole block — roughly twelve seconds on the network studied. The maker has two defences: shorten the time its quote is exposed, or screen the flow and decline orders most likely to be informed. A permissionless pool must fill every order that clears its invariant, so only the first lever is open to it; a curated maker can use both.
The data. The study benchmarks four USDC-WETH pools on Uniswap v2 and v3 against the Binance ETH-USDT best bid and ask, reduced to a one-second midpoint, from 1 June to 29 August 2026. After removing dust, the sample holds 1,725,539 swaps; the largest pool, Uniswap v3 with a 1 bp fee, accounts for 1,290,417 of them. The netting statistic uses the full Binance trade tape of 260 million trades.
The headline numbers. Price moves on Binance are far from bell-shaped: in 73% of one-second intervals the midpoint does not move at all, and the excess kurtosis is 154. Going from a twelve-second to a one-second update removes 84.5% of the expected price shock. A maker quoting every twelve seconds therefore has to turn away no less than 84.5% of the informed orders to do as well as a venue that updates every second — a normal-distribution model would put the bar at 71.1%. The author adds that this does not depend on knowing exactly how toxic the flow is: across the informed shares allowed by the tightest bound, the threshold stays between 84.5% and 85.2%, or in the post's words, roughly six in seven informed orders.
What happens in trends. When the market trends, the stale pool is picked off at execution: fills that go with the trend obtain an effective spread 11.1 bps lower than fills against it, with the same sign in all four pools. Using missed proposer slots, the study finds that the gap a stale pool must absorb grows with elapsed staleness at an elasticity of 0.45, close to the diffusion rate of one half. Predictable drift, it concludes, is a centring problem: it widens a quote anchored on the current price, but its cost disappears once the quote is centred on the expected reference.
Why it is worth reading. The model's threshold is stated to be scale-free under a driftless diffusion, depending only on the ratio of the two time horizons and the informed share, not on volatility, fee or trade size; and the author argues that the shape of the tail, not just its width, is what the maker prices. For a chain debating faster blocks and for pools weighing curated order flow, the post puts both levers on one scale, measured on mainnet data rather than assumed.
Source: ethresear.ch — https://ethresear.ch/t/how-much-screening-buys-a-second-speed-selection-and-the-makers-spread/26158