Latency & High-Frequency Trading

In a market that matches orders continuously, being faster than the next participant is worth money, sometimes a great deal of it. Latency is the time between an event (a price change, a signal) and your order reaching the matching engine, and the competition to minimize it has reshaped market structure: microwave towers between Chicago and New Jersey, co-located servers, FPGA order gateways, and firms that spend fortunes to shave nanoseconds. This concept explains why speed pays, where exactly the edge comes from, and the influential argument that the whole race is an artifact of market design.

Why speed matters: two distinct edges

Speed is valuable for two mechanically different reasons.

1. Winning the queue (liquidity provision). By Order Book Mechanics price–time priority, the first order at a price level fills first. In a one-tick market where the spread cannot narrow, time priority is the entire competition: front-of-queue orders fill against benign flow and capture the spread, while back-of-queue orders fill only when the queue is being swept because the price is about to move, i.e., they are Adverse Selection-poisoned. Being microseconds faster to post (and to re-post after a cancel) buys a better queue position, which is a direct increase in realized spread capture. This is the market-maker's speed edge, and it is why The Avellaneda-Stoikov Model-style quoting in practice is inseparable from a latency race.

2. Sniping stale quotes (liquidity taking / latency arbitrage). When new information arrives, a move in a correlated instrument, an index future ticking, resting quotes are momentarily stale. The fastest trader can pick off ("snipe") a quote at the old price before the maker cancels it. The classic setup is two venues or two correlated assets: the price of one moves, and for a few microseconds a limit order on the other is mispriced. Whoever reacts first takes the free money; whoever is a maker there wants to cancel first. This is latency arbitrage, and it is a pure speed rent.

The sniping model and its consequence

Budish, Cramton & Shim (2015) formalize the second edge and draw a sharp conclusion. Consider a market maker quoting an asset whose "true" value jumps by $J$ on public news at Poisson rate. When a jump happens, the maker wants to cancel the now-stale quote; a swarm of snipers wants to hit it. If the maker is not the very fastest, they get sniped with some probability and lose about $J$ each time.

Their central result: this creates a mechanical, built-in adverse-selection cost that competition cannot eliminate. The expected sniping loss per unit time is roughly

$\text{loss} \sim \Pr(\text{sniped} \mid \text{jump}) \times (\text{jump rate}) \times J,$

and because someone is always the fastest, the maker must widen the spread to cover it. Crucially, the spread does not shrink to zero even with unlimited competition among makers, because the binding cost is being last to react to public information, a cost that more competition does not remove, it just redistributes to whoever is fastest. The rent from being fastest is a constant that persists no matter how many firms compete; hence the arms race, everyone must keep spending on speed just to not be the slowest, a prisoner's-dilemma expenditure that produces no social value and is ultimately paid by liquidity takers through wider spreads.

The proposed fix: frequent batch auctions

Budish et al.'s design implication is elegant: the arms race is a symptom of continuous-time trading, where the tie-break for two orders arriving "at the same time" is decided by nanoseconds. Replace continuous matching with frequent batch auctions, collect orders over a short window (say 100 ms) and clear them all at a single uniform price. Within a batch, orders are ranked by price, not time, so being a microsecond faster no longer wins; the speed advantage collapses to the batch interval. This converts a competition on speed into a competition on price, which is the socially useful one. The argument is one of the most cited in market-design and underlies real experiments in batch-auction venues.

Worked example

A market maker quotes a stock; correlated-index news causes the fair value to jump by J = \0.02 at rate 50 times per hour. The maker's cancel reaches the exchange in 300 microseconds; the fastest sniper reaches it in 100 microseconds. On each jump the sniper picks off the stale quote before the cancel lands, so the maker loses \0.02 per jump: 50 \times \0.02 = $1.00$ per hour per quoted unit, purely from being 200 microseconds slower. To break even the maker must earn that back from uninformed flow, widening the spread. Now the maker invests in a faster gateway that beats the sniper, the loss vanishes, but the sniper firm then invests to get faster still. Neither firm captures a durable edge; the exchange and hardware vendors do, and the spread stays wide enough to fund the race. Under a 100 ms batch auction, the jump would be reflected in the next clearing price and the sniping opportunity would essentially disappear.

Failure modes and caveats

• Speed is not alpha. A latency edge decays to zero the moment a competitor matches it; it is a Red Queen's race requiring perpetual capex. It is also fragile, a single slow link, a co-lo outage, or a matching-engine change can erase it.
• HFT is not one thing. "HFT" spans benign electronic market-making (which narrows spreads and adds liquidity, per Menkveld) and predatory latency arbitrage (which taxes it). Conflating them muddies the policy debate.
• Liquidity can be illusory. Ultra-fast quotes can be cancelled faster than slow participants can hit them; displayed depth from HFTs may evaporate exactly when it is needed (the "phantom liquidity" concern raised after flash crashes).
• Winner-take-all fragility. Because only the fastest wins each race, small latency differences produce large payoff differences, a highly nonlinear, unstable competition prone to overinvestment.
• Market-design dependence. Everything here is contingent on continuous-time price–time priority; change the matching rule (batch auctions, speed bumps, randomized delays) and the entire edge structure changes.

In interviews

Be able to separate the two speed edges: winning queue position for spread capture (a liquidity-provision edge rooted in Order Book Mechanics price–time priority) versus sniping stale quotes on news (latency arbitrage, a liquidity-taking edge). Explain why the race persists: because someone is always fastest, being last to react to public information is an unavoidable Adverse Selection cost that competition redistributes rather than removes, so makers must widen spreads and everyone must keep spending on speed, the Budish–Cramton–Shim arms-race argument. The sophisticated close is the design fix: frequent batch auctions turn a competition on speed into a competition on price by ranking within a batch by price, not time. A good candidate also notes the nuance that electronic market-making can genuinely tighten spreads even as latency arbitrage taxes them.