Latency & the Speed Arms Race

You have spent five lessons learning what to quote — spreads that cover adverse selection, skews that bleed off inventory, the math of Avellaneda–Stoikov. This lesson is about how fast you can react after you quote it. Because a market maker’s quote is a standing promise to trade at a price, and the world does not wait for you. The instant new public information moves the fair value, your resting quote is stale — too cheap or too dear — and somewhere a faster trader is already racing to pick it off before your cancel arrives. For a maker, slowness is not an inconvenience; it is a tax, paid in toxic fills, every microsecond you lag the fastest counterparty. This is the lesson where milliseconds turn into money, and money turns into an arms race.

Slowness is a tax: the stale-quote problem

Analogy. Imagine you run a currency-exchange kiosk and you post your rates on a big board. A bulletin flashes across every TV in the airport: the euro just jumped. For a few seconds your board still shows the old rate. Anyone who sprints to your window in those seconds buys euros from you at yesterday’s price and walks straight to the kiosk next door to sell at the new one. You haven’t been outsmarted — the news was on every screen — you’ve just been out-sprinted. Your only defense is to update the board faster than they can run to your window.

Definition. A quote is stale when the fair value has moved but your resting order is still at the old price. Every resting quote carries this risk continuously: the maker is short an option. By posting a firm price you have effectively written a free option to the rest of the market — the right, but not the obligation, to trade against you at your stale price the moment it becomes favorable to them. The faster the world can exercise that option versus how fast you can cancel it, the more it costs you.

Worked example. Suppose fair value jumps from $100.00$ to $100.20$ and your offer sits at $100.05$. The fill costs you the difference between what you sold for and what it’s now worth:

On a 1,000-share fill that’s a $150 hit — on a single quote, from a single public print, with zero informational disadvantage on your part. If your tick-to-trade reaction time were short enough to cancel before the sniper arrived, that $150 stays in your pocket. Multiply by thousands of such events per day across thousands of symbols and you see why speed is worth nine-figure budgets.

Misconception. “Adverse selection is only about insiders / private information.” No. The Glosten–Milgrom lesson covered the informed-trader flavor. Stale-quote sniping is a second, distinct flavor: the information is fully public, and the only edge the winner has is that they processed and acted on it faster. You can have zero informed traders in your flow and still bleed to speed.

Stale-quote sniping & latency arbitrage

Analogy. Two people see the same lottery number announced on TV at the same instant. The prize goes entirely to whoever physically reaches the claim office first. The knowledge was identical and free; the race decided everything. That is latency arbitrage.

Definition. Stale-quote sniping (a.k.a. latency arbitrage) is the strategy of detecting that a resting quote has become mispriced relative to new public information and executing against it before the maker can cancel. Concretely: a signal arrives (a print on a correlated instrument, a futures tick, an index move). Every fast participant sees it at nearly the same time. The maker wants to cancel the now-stale quote; the snipers want to take it. It is a pure foot-race between one cancel message and many take messages, all triggered by the same public event.

Why it’s adverse selection by speed. The sniper is not better informed — they read the same public tape. They are better positioned and faster. The fills the maker loses are systematically the ones where the maker was about to be right that the price moved; that is exactly what makes the flow toxic.

Budish–Cramton–Shim: the CLOB makes competition a speed race

Analogy. Picture an auction where, instead of everyone bidding once, the auctioneer accepts bids continuously and the very first hand up wins. Bidders stop competing on price and start competing on reflexes — hiring sprinters, standing closer, watching the auctioneer’s lips. The auction format itself manufactured a race that has nothing to do with who values the item most.

The argument. In their influential 2015 paper “The High-Frequency Trading Arms Race: Frequent Batch Auctions as a Market Design Response,” Eric Budish, Peter Cramton, and John Shim make three linked claims:

1. The continuous limit order book (CLOB) is the culprit. Because the book processes messages one-at-a-time in continuous time, whenever public information moves, it creates a brief mechanical arbitrage: a sniping opportunity that is won purely on latency. This recurs constantly — it is “baked into” continuous trading, not a glitch.
2. The result is an arms race that is a prisoners’ dilemma. Each firm must invest in speed because rivals do; the spending is privately rational but socially wasteful. It does not improve price discovery — the prices were already moving — it just transfers the same rents to whoever is fastest, while burning real resources (microwave towers, FPGAs, fiber) in the process.
3. Frequent batch auctions (FBAs) are the proposed fix. Replace continuous matching with very frequent discrete-time auctions — e.g. batch all orders arriving within each, say, 100-millisecond window and clear them together at a single uniform price. Within a batch, time priority disappears: being a nanosecond faster no longer wins, because everything in the window is treated as simultaneous. That removes the speed advantage and turns competition back toward price.

Worked intuition. In a CLOB, if a signal arrives and you are 1 microsecond faster than the next firm, you win the entire opportunity. In an FBA with a 100 ms batch, both of you land in the same batch, so your 1 µs edge buys you nothing — you compete on the price you’re willing to quote instead. The mechanical-arbitrage rent collapses.

Trade-off / pitfall. Batching is not free: it adds a small delay to every trade (worst case one batch interval), and choosing the interval is delicate — too long and you hurt genuine price discovery and impatient traders; too short and snipers still win. And no major equity exchange has adopted pure FBAs, so it remains a celebrated theory and partial inspiration for speed bumps, not the status quo.

Anatomy of latency: the budget

Analogy. Shipping a package across the country has a travel time (distance ÷ vehicle speed) and a handling time (sorting, loading, customs). Latency is the same: propagation is the travel, processing is the handling, and you optimize both. You can buy a faster truck (better medium) and a faster warehouse (better hardware).

The components.

• Propagation delay — how long the signal takes to physically travel the distance. Bounded by the speed of light in the medium. Light in vacuum is $c \approx 299{,}792$ km/s; light in glass fiber travels at roughly $\tfrac{2}{3}c \approx 200{,}000$ km/s (the refractive index of fiber is ~1.5); microwave through air travels at very nearly $c$.
• Processing delay — the time inside your own systems from “data in” to “order out.” This is the tick-to-trade path: receive the market-data packet, decode it, run the strategy logic, encode the order, push it onto the wire.
• Colocation — renting rack space inside the exchange’s data center so your propagation to the matching engine is a few meters of cable, not miles. Table stakes for any serious maker.

The processing ladder (fastest wins):

Worked example — propagation from distance. The Chicago futures market (CME, Aurora IL) and the New Jersey equity venues are about 1,300 km apart one-way. Compute the one-way propagation floor:

• Speed-of-light vacuum floor:

• Through fiber (light at ~$\tfrac{2}{3}c \approx 200{,}000$ km/s), and real fiber routes wind rather than go straight — the deployed path is more like 1,300 km of cable but at the slower medium:

• Through microwave (~$c$, straighter line-of-sight path): roughly 4.5 ms one-way in practice.

So switching the Chicago↔NJ link from fiber to microwave shaves on the order of ~2 ms off the one-way trip — an eternity in this business.

Microwave & the Chicago↔NJ route

Analogy. Fiber is a winding mountain road; microwave is a hang-glider flying straight over the peaks. The glider’s path is shorter (a straight line beats a road that follows terrain) and it moves through a faster medium (air ≈ c vs glass ≈ ⅔c). Two advantages stack.

Why microwave beat fiber on this route. Two compounding reasons:

1. Faster medium. Microwaves through air travel at ~$c$; light in fiber at ~$\tfrac{2}{3}c$. Same distance, ~50% faster signal.
2. Straighter path. Fiber must follow rights-of-way — railways, highways, conduits — so the cable is longer than the crow-flies distance. A microwave network is a chain of line-of-sight towers placed as close to a straight line between Chicago and New Jersey as geography and tower spacing allow.

Together these cut the one-way time from ~6.5 ms (fiber) toward the ~4.5 ms region — close to the vacuum floor of ~4.3 ms.

Trade-off / pitfall. Microwave is lower bandwidth and weather-sensitive: heavy rain, snow, and fog attenuate the signal (“rain fade”), and links can drop. Firms therefore run microwave for the latency-critical trickle (a few key signals) and keep fiber as a high-bandwidth, all-weather backup. Millimeter-wave and even laser/free-space optics push further but are even more weather-fragile. The economics are brutal: a tower chain costs a fortune and the whole advantage can be one or two milliseconds — which only pays if you can convert that lead into winning races.

Winner-take-all economics

Analogy. A 100-meter Olympic final pays gold to the runner who wins by one thousandth of a second; silver gets a medal, but in a pure latency race there is no silver — second place gets nothing on that opportunity. So everyone trains insanely hard for a margin no spectator can even see.

The economics. In a pure latency race, each fleeting opportunity is winner-take-all: only the single fastest firm captures it; everyone else’s order arrives to find the quote already gone. This has three consequences:

1. Huge incentive to spend on tiny edges. Even a sub-microsecond lead, if it converts marginal races from losses to wins, can be worth far more than its cost — so firms rationally pour capital into shaving nanoseconds.
2. Arms-race rents are socially wasteful. The resources spent on speed (towers, FPGAs, exotic links) mostly redistribute existing rents to the fastest firm rather than create new value — price discovery would have happened anyway. Budish–Cramton–Shim call this the prisoners’-dilemma waste.
3. Industry concentration. Because second place earns nothing in the latency race, the winner-take-all structure pushes the industry toward a handful of firms that can afford to be at the very top of the speed distribution.

Worked intuition. Suppose an opportunity is worth $10 and occurs 1,000 times a day. The fastest firm captures (nearly) all $10,000/day; the second-fastest captures almost $0 of it. So a firm will rationally spend up to ~$10,000/day-equivalent to move from #2 to #1 — even though the social value created by that spend is roughly zero. That gap between private and social return is the arms-race pathology.

Speed bumps & market-design responses

Analogy. If sprinting to the window is the problem, one fix is to make everyone pause for the same fraction of a second at the door — long enough that the maker can pull a stale sign before the sprinter reaches the counter. A uniform delay neutralizes the foot-race without changing who’s allowed in.

The responses.

• IEX 350 µs speed bump — the “magic shoebox.” IEX coils 38 miles of fiber in a box, adding a fixed 350-microsecond delay to inbound orders. That tiny delay is enough for IEX’s pricing logic to update its reference (the “crumbling quote” signal) and reprice/protect resting orders before a latency arbitrageur’s order can pick them off. It blunts sniping while treating all incoming orders the same.
• Asymmetric speed bumps — delay only aggressive (liquidity-taking) orders while letting makers cancel without the delay. This explicitly hands the cancel race to the maker, gutting latency arbitrage — but critics argue it’s a thumb on the scale that lets makers post quotes they can yank at will (“last look”–flavored).
• Frequent batch auctions — the Budish–Cramton–Shim discrete-time fix from earlier: batch a short window and clear at one price so simultaneity, not speed, governs the match.

Controversy. Speed bumps are genuinely divisive. Supporters say they neutralize a socially wasteful arms race and protect resting liquidity. Detractors say they (a) fragment the market with non-uniform timing, (b) the asymmetric kind can advantage makers unfairly, and (c) they don’t eliminate the speed incentive so much as move the goalposts. Regulators have approved some (IEX became a registered exchange in 2016) while scrutinizing others.

The maker’s cost-benefit: is speed worth it?

Analogy. Buying speed is like buying insurance against being pickpocketed: you pay a fixed premium (capex + opex) and in return you lose fewer wallets (toxic fills). It’s worth it only if the wallets you save exceed the premium — and it does nothing about getting mugged (informed traders) or about carrying too much cash (inventory risk).

The calculation. A speed investment is justified when the reduction in sniping/adverse-selection losses plus the value of higher cancel-success exceeds the amortized cost of the upgrade:

Worked cost-benefit table. Suppose a maker today loses $40,000/day to stale-quote sniping. A microwave + FPGA upgrade costs $30M capex (amortized over ~3 years ≈ 1,000 trading days → $30,000/day) plus $5,000/day opex, and it cuts sniping losses by 70%.

At these numbers the upgrade loses $7,000/day — speed is not automatically worth it. Flip one input (say sniping losses are $70,000/day, cut 70% → $49,000 saved) and it flips to +$14,000/day net. The decision is entirely about whether your toxic-fill bleed is large enough to amortize the hardware.

Misconception / pitfall. “If I’m the fastest, I’ve solved market making.” Speed is necessary but not sufficient. It defends against latency adverse selection only. It does nothing for:

• Informed adverse selection — a counterparty with genuine private information beats you regardless of your latency (Glosten–Milgrom).
• Inventory risk — being fast doesn’t stop you from accumulating a dangerous long/short position; you still need skew and inventory control.

Speed is one leg of the stool. Kick out spread-setting, skewing, and toxicity filtering and the fastest firm on earth still goes broke.

Key Takeaways