Making markets
High-frequency trading
High-frequency trading is less a strategy than a speed at which strategies run: market making that requotes in microseconds, and arbitrage that trades on a price change in one market before slower participants have updated in another. The edge is measured in microseconds and nanoseconds, and is paid for in hardware, networks and engineering.
At a glance
At a glance
- Makes money from
- Tiny, very frequent gains from spreads and from reacting to price changes first
- Holds positions for
- Milliseconds to seconds; usually flat by the end of the day
- Trades
- Futures, shares, currencies and government bonds on electronic venues
- Biggest risk
- Being slower than the competition – then yours are the stale quotes that get picked off
- Who does it
- Firms such as Hudson River Trading, Jump Trading, Virtu and Tower Research
- 537 – latency races per FTSE 100 stock a day (LSE data, 2015)
- 5–10µs – typical winning margin in a race (FCA study)
- 22% – of FTSE 100 volume traded in races (FCA study)
- 3.93ms – light-speed floor, Chicago to New Jersey (one way)
- 350µs – the IEX exchange’s speed bump (38 miles of coiled fibre)
- 13.9ns – record hardware reaction in a benchmark (2024, network path only)
The business
When the S&P 500 futures tick up in Chicago, the same news reaches every related market – the index ETF, the shares in the index, options – a few milliseconds later. Whoever sees it first can buy from the quotes that have not yet moved, or move their own quotes before anyone trades against them. The profit on each event is small; there are millions of events.
The costs are large and mostly fixed: servers in the exchanges’ own data centres (co-location), direct market-data feeds, private microwave and millimetre-wave links between trading centres, custom hardware and the engineers who build it. That favours a small number of firms at the frontier, and makes the business strongly winner-take-most: in the UK’s best-studied data, the top three firms won about 55% of races and the top six about 82%.
Latency arbitrage between two markets
A race, measured
In 2020 economists at the UK’s Financial Conduct Authority and the University of Chicago studied every message sent to the London Stock Exchange over nine weeks, and found the races directly: bursts of messages from several firms, within microseconds, all trying to trade or cancel at the same price.
What the London Stock Exchange data showed
| Measure | Finding |
|---|---|
| How often | 537 races per FTSE 100 stock per day – about one a minute |
| How close | The winner typically beats the first loser by 5–10 microseconds |
| How much | About 22% of FTSE 100 volume, and a third of the effective spread |
| What it is worth | About half a tick, roughly £2, per race |
| The cost to investors | A “latency tax” of 0.42 basis points on all trading |
| Who wins | Top 3 firms about 55% of races; top 6 about 82% |
| Without races | The cost of liquidity would fall by about 17% |
Worked example
How big is the prize?
Using the study’s figures for the FTSE 100 alone, and 252 trading days.
- Races a day stocks53,700
- Value at stake a day£107,400
- A year≈ £27m
About £27 million a year from the FTSE 100, shared among a handful of firms – the study estimated about £60 million across UK equities and about $5 billion a year globally. Small per race, small even in total next to the cost of the arms race, which is why only the fastest few find it worth fighting for.
How long things take
The physics that matters
Light in a vacuum covers about 300 kilometres a millisecond; in glass fibre it is about a third slower. Chicago to the New Jersey data centres is about 1,180 km in a straight line, so the floor is about 3.9 ms one way through air, and about 5.8 ms through a perfectly straight fibre. Real fibre follows roads and railways: the dedicated Spread Networks cable, finished in 2010 at a reported cost of hundreds of millions of dollars, did the round trip in about 13 ms. Microwave towers, which go straight and through air, cut that to about 8 ms – at the price of losing the signal in heavy rain.
Worked example
The speed-of-light limit
Chicago to New Jersey, about 1,180 km in a straight line.
- Through air km a second3.94 ms
- Through straight fibre, glass’s refractive index5.8 ms
- The real fibre routeLonger than a straight line≈ 6.5 ms
Microwave towers get within a few per cent of the 3.9 ms floor; fibre cannot get below 5.8. A millisecond and a half is an eternity in a race decided by 5 microseconds – the reason firms paid to build tower networks.
Economists Eric Budish, Peter Cramton and John Shim showed in 2015 that the race is mechanical. The prices of the S&P 500 future and the matching ETF are almost perfectly correlated over minutes but not over milliseconds, which creates a fresh arbitrage every time one moves. The median window to trade it shrank from 97 ms in 2005 to 7 ms in 2011, but the profit on each opportunity did not fall: speed competition only moved the prize to whoever was fastest. Their proposed fix is frequent batch auctions – matching orders every fraction of a second rather than continuously.
How it trades
- Short-horizon signals: order-book imbalance, trades on related instruments, the last few price changes.
- Market making at high speed: requote the moment fair value moves, so your quotes are rarely the stale ones.
- Cross-venue and cross-product arbitrage: the same exposure priced slightly differently in two places.
- Everything engineered for latency: network cards that bypass the operating system, pinned processor cores, code with no memory allocation or locks between the market data and the order, and FPGAs for the most time-critical steps.
Who wins the races
Risk and debate
Whether the speed race helps markets is argued over. High-frequency market makers have narrowed spreads for ordinary investors; the races they run are a cost those investors pay. Speed bumps such as the 350-microsecond delay at the IEX exchange exist to blunt them, and regulators have accepted that delays under a millisecond can be fair.
The Flash Crash · 6 May 2010
High-frequency firms passed contracts back and forth as one large sell order hit the E-mini future, and US prices fell and recovered within minutes – the episode that put high-frequency trading under scrutiny.
A straighter fibre · August 2010
Spread Networks opened its dedicated Chicago–New Jersey cable, cutting the round trip to about 13 ms. Within a few years microwave networks had beaten it, at about 8 ms.
An exchange with a speed bump · June 2016
The SEC approved IEX as an exchange, with a 350-microsecond delay built from 38 miles of coiled fibre, and said delays under a millisecond could be acceptable.
The races measured · January 2020
The FCA published the study of London Stock Exchange messages above, later in the Quarterly Journal of Economics.
A 13.9-nanosecond benchmark · June 2024
A hardware vendor set a record of 13.9 ns in the STAC-T0 benchmark – the time from market data in to order out on the network path alone, with almost no trading logic.
Who works here
Quant trader or researcher
Builds the short-horizon signals and fill models, studies which races are worth entering and at what cost, and tunes parameters as competition and markets change. Works on message-level data where a microsecond matters.
Software engineer
Writes and measures the code on the path from market data to order: feed handlers, the order gateway, inline risk checks. Profiles latency with hardware timestamps, and cares about the worst case (jitter) as much as the average.
Hardware and network engineer
Designs FPGA logic for feed handling and pre-set orders, and the networks between data centres, including microwave and laser links.
In interviews for this desk
Engineering interviews dominate: C++, data structures, systems and networking, and how to shave latency. Researchers are asked about statistics on short horizons and noisy data; traders about probability and quick reasoning.
What is the fastest a signal can travel from Chicago to New Jersey?
About 1,180 km at about 300,000 km a second: 3.9 ms through air. Through a straight fibre, about 1.47 times slower, 5.8 ms – and the real fibre route was about 6.5 ms, which is why firms built microwave links.
Design the path from a market-data packet arriving to an order leaving.
A network card that bypasses the kernel, or an FPGA feed handler; a busy-polling thread pinned to an isolated core; parse the packet in place; lock-free single-producer queues between stages; pre-allocated order objects; no system calls, allocation or logging on the path; inline pre-trade risk checks and a kill switch; hardware timestamps to measure it all.
Your system’s average latency is excellent but its 99.9th percentile is poor. Why does that matter, and what causes it?
Races are decided at the moment something happens, often when the market is busiest – exactly when slow outliers occur. Causes include cache misses, page faults, garbage collection or allocation, interrupts on the trading core, and logging; the fixes are isolation, pre-allocation and moving work off the critical path.
Exchanges send each market-data feed twice, on A and B lines. What do you do with them?
Arbitrate: take each message from whichever line delivers it first, using sequence numbers to drop the duplicate and detect gaps. It cuts latency and protects against a line dropping packets.
What would frequent batch auctions change?
Matching orders in batches every fraction of a second, rather than continuously, turns a race decided by microseconds into competition on price within the batch. The prize for being a few microseconds faster disappears; Budish, Cramton and Shim argue that would cut trading costs.
Why does your place in the queue matter to a high-speed market maker?
At the front of the queue you are filled by small, mostly uninformed orders; at the back, mainly when a large informed order sweeps the level. Getting to the front quickly after a price change is worth both more fills and better ones.
Firm guides: Hudson River Trading · Jump Trading · Citadel Securities
Glossary
- Latency
- The time from an event to the response to it.
- Latency arbitrage
- Trading against quotes that have not yet reacted to public news.
- Co-location
- Servers placed inside the exchange’s own data centre.
- Kernel bypass
- Network input and output that skips the operating system.
- FPGA
- A reprogrammable chip used to process market data and send orders in hardware.
- Jitter
- Variation in latency – the slow outliers.
- Tick-to-trade
- The time from a market-data update arriving to the order it triggers leaving.
- Speed bump
- A deliberate delay applied to incoming orders.
- Frequent batch auction
- Matching orders in batches at short fixed intervals instead of continuously.
- A/B feeds
- Two copies of an exchange’s market data, sent on separate lines.
Further reading
- Quantifying the High-Frequency Trading “Arms Race” – Matteo Aquilina, Eric Budish and Peter O’Neill, 2022
- The High-Frequency Trading Arms Race: Frequent Batch Auctions as a Market Design Response – Eric Budish, Peter Cramton and John Shim, 2015
- Information Transmission between Financial Markets in Chicago and New York – Gregory Laughlin, Anthony Aguirre and Joseph Grundfest, 2014
- Algorithmic and High-Frequency Trading – Álvaro Cartea, Sebastian Jaimungal and José Penalva, 2015