The beams are infrared, which means you can’t see them, but lasers are now flashing stock-market data through the skies over New Jersey. If they work well there, they might soon be flashing over London too. Lasers are the latest tool for high-frequency trading: the fast, entirely automated trading of large numbers of shares and other financial instruments.
Originally, the data needed for high-frequency trading travelled almost exclusively via fibre-optic cables, in which signals move at about two-thirds of the speed that light travels in a vacuum. It’s a tried, trusted technology. Fibre has the bandwidth to transmit the huge volumes of data spewed out by today’s financial markets. Although accidents do happen (farmers, for example, sometimes cut the buried cables when ploughing), fibre-optic links are very reliable. But 125,000 miles per second isn’t fast enough if other market participants are faster. So there’s a race on. The theory of relativity tells us that nothing can travel faster than light in a vacuum. Air, unlike glass fibre, slows light down almost imperceptibly. Sending light or radio signals through the atmosphere is less than a tenth of 1 per cent slower than the speed of light in a vacuum.
Every way of sending market data through the air has its limitations. Wireless microwave transmission, using ‘dishes’ (antennae) on a series of towers, is an old technology, but improved versions get you the necessary speed. Gales can blow the dishes out of alignment, and make it unsafe for workers to climb up to realign them, but those involved tell me a good microwave link will work well more than 99 per cent of the time. Its bandwidth, however, is limited, meaning ‘you have to pare down the data,’ as one user put it, removing what isn’t necessary. Specialised computer hardware can do the necessary editing, compression and decompression in a couple of microseconds (millionths of a second), but we have now reached the point where delays measured even in nanoseconds (billionths of a second) matter.
Higher-frequency millimetre waves provide greater bandwidth than microwaves, so there’s less need for ‘editing’, but are more easily disrupted by bad weather. Three years ago, millimetre-wave links ‘would go down when someone sneezed’, says Michael Persico, whose Chicago firm, Anova Technologies, uses these links. Even today, they tend to stop working in heavy rain. So Anova has begun to use lasers to supplement its millimetre waves. Lasers have vulnerabilities too (fog is a big problem), but Anova says its test results show that a combination of lasers and millimetre waves is about as reliable as fibre-optic cable, while getting close to the holy grail of the speed of light in a vacuum.
Not so long ago there was a vogue for asserting that globalisation had ‘ended geography’ and created a ‘flat world’. When financial trading was a matter of human beings looking at screens, that had a certain plausibility, because the intrinsic slowness of humans’ eyes and brains easily masked the small disparities in the time taken to transfer data between different geographic locations. But now that computers, not humans, are doing the trading, geography matters exquisitely. With any of these technologies – fibre-optic cable, microwave, millimetre wave, laser transmission through the atmosphere – the exact route taken is crucial.
The shortest and therefore fastest route on the surface of the earth between any two places is called a ‘geodesic’ or great circle. When you talk to high-frequency traders (something I’ve done a lot over the past four years), you quickly learn that the world’s financially most important geodesic – the spinal cord of US capitalism – runs from Aurora, a town in Illinois that’s now essentially an outer suburb of Chicago, to northern New Jersey. It’s now close to impossible, I’m told, to put new microwave dishes along that crowded geodesic: all the space on the towers closest to it is already rented. The dishes can interfere with one another, so you need permission from the Federal Communications Commission to build new towers or install new dishes. As a result, no one is easily going to beat McKay Brothers, which owns the fastest, geodesic-hugging Aurora-New Jersey microwave link.
Aurora matters to global finance because in 2012 the Merc, the Chicago Mercantile Exchange, relocated its electronic trading system to a new data centre there. The Merc trades futures: originally, futures on eggs, onions and other agricultural commodities, but since 1972 financial futures as well. Chicago futures trading used to be done face to face (by voice, or eye contact and hand signal) in raucous, crowded trading pits. The Merc’s pit traders fiercely resisted the coming of electronic trading: its leading advocate, Leo Melamed, received frequent death threats. By 2004, however, the resistance had crumbled, and now nearly all the Merc’s trading is electronic.
The Merc’s first fully electronic product was the E-Mini financial future, launched in 1997. It tracks the S&P 500 index, made up of the five hundred leading US stocks. The buyer and the seller of an E-Mini each maintain a ‘margin’ deposit on account at the Merc’s clearing house. Each night, the clearing house adjusts those deposits. If the S&P 500 index has risen by a single point, $50 is transferred from the seller’s account to the buyer’s; if it has fallen by ten points, say, $500 shifts from the buyer to the seller. If their deal is for a thousand E-Minis, the latter sum becomes $500,000.
Traders tell me that new information relevant to the overall value of US shares almost always shows up first in orders for and in the prices of the E-Mini, and a fraction of a second later in the underlying shares; financial economists have documented the same pattern. The reason for it is probably that the E-Mini gives you greater ‘leverage’: a modest ‘margin’ deposit permits gains (and, of course, also losses) corresponding to buying or selling a large and expensive block of shares. So if traders think they or their automated trading systems have an information edge, it is to the E-Mini they will usually turn first. The big crises of modern US stock markets have tended to show up first in the E-Mini (or, before 1997, in its predecessor, the S&P 500 pit-traded future) and a little later in the stock market.
Changes in the electronic order book for the E-Mini are crucial information for automated share trading (this particular game is now too fast for human players, who wouldn’t be able to react quickly enough to the changes). Suppose the price of the E-Mini has fallen, or even simply that the number of offers (sell orders) has risen sharply and the number of bids (buy orders) has fallen. Over the next fraction of a second, falls in the prices of the underlying shares are more likely than increases. I’m told that a similar pattern generally holds for US Treasury bonds, with the prices of the bond futures traded in Aurora leading changes in the prices of the underlying bonds (although sometimes it’s the other way around). Foreign exchange is different. Although Aurora is the world’s main site of currency futures trading, the dominant foreign exchange markets are still the ‘spot’ markets in which foreign exchange for immediate delivery is traded among banks, hedge funds and so on.
US shares, US government bonds and much of ‘spot’ foreign exchange are traded in data centres in northern New Jersey. These data centres resemble giant warehouses, and their size is the reason the operation has shifted from its traditional sites in Manhattan to townships in New Jersey no tourist has ever heard of: real estate is much cheaper there. Data centres often have high-security features, including a two-door entrance like a spaceship airlock. They’re often windowless, and sometimes freezing cold: fierce air conditioning is needed to extract the heat generated by the tens of thousands of computer servers they contain. (The small number of onsite maintenance workers stay in warm rooms unless something goes wrong or new equipment is being installed.) Data centres are huge consumers of electricity, and the combination of so many computers and all that air conditioning makes for a lot of noise. Andrew Blum, author of a fine book on the physical reality of the internet, describes visiting a data centre as ‘like stepping into a machine … as loud as a rushing highway’.
There are four main share-trading data centres in the US. The New York Stock Exchange owns its own data centre in Mahwah, New Jersey. Nasdaq’s is in Carteret. The computer systems of their biggest rival, BATS (Better Alternative Trading System, set up in 2005 by a team from the Kansas City high-frequency trading firm Tradebot), are currently in NJ2 in Weehawken, owned by the data centre managers Savvis. The share-trading systems of the fourth main exchange in the US, Direct Edge (recently bought by BATS), are in NY4, a giant multi-user data centre owned by Equinix, the largest global provider of these structures. Despite its name, NY4 is in Secaucus, New Jersey, and much of the trading of bonds, foreign exchange and options also takes place there.
Each data centre contains the exchange’s or other trading venue’s matching engines, which are the computer systems that maintain its electronic order books, search for a match (a bid and an offer at the same price for the same share, bond or currency) and, if they find one, consummate the trade and generate electronic ‘confirms’ to tell the parties to the trade that it has taken place. Surrounding the matching engines are the order gateways, computer systems that receive the electronic messages containing traders’ orders and cancellations of orders, then send those messages to the correct matching engine, and dispatch ‘confirms’. An exchange will also have at least one ‘feed server’, which collates information on every change in the exchange’s order books and sends it out in a continuous stream of messages referred to as the ‘raw’ data feed.
Nowadays, the majority of orders for US and European shares and futures (and increasing proportions of the orders for bonds and foreign exchange) are generated by computer systems which are themselves often located in the data centre that processes them. High-frequency trading firms – and also, for example, the big banks that act as brokers for institutional investors – pay the exchanges or other data centre owners for space in which to place their computer servers. I’m told that a ‘rack’, a cabinet-sized space that can accommodate thirty or forty computers, costs between $1500 and $15,000 a month. A big trading firm will need multiple racks in each data centre, or even a whole ‘cage’ (to protect against rivals’ tampering, trading servers are often kept in locked cages), which can easily cost more than $1 million a year. The owners of data centres can charge these high rents because it’s a very concentrated world. Most global financial trading probably takes place in no more than 15 data centres: six in the US (the four big share-trading centres, the Merc’s in Aurora, and a big multi-user centre closer to the Chicago Loop); another five in Europe; and a handful in the rest of the world.
An institutional investor will set a goal for the computer programs of a bank executing orders on its behalf, for example to buy a hundred thousand Apple shares. If the programs immediately and visibly placed the whole order in the electronic order books of the exchanges on which Apple shares are traded, it would drive up their price sharply. So the bank’s programs (its ‘smart order router’ and its ‘execution algorithms’) split the ‘parent’ order into up to a thousand ‘child’ orders and execute them as surreptitiously as possible, often trying to make them look like large numbers of independent orders from small investors. The order router will usually at first send child orders to ‘dark pools’ (which are not exchanges, but trading venues in which participants, whether human or algorithmic, cannot see the electronic order books). If the router then has to send child orders to exchanges such as BATS, Direct Edge, Nasdaq or the New York Stock Exchange – in which order books are visible – it will keep them small and therefore hopefully inconspicuous, buying or selling as few as a hundred Apple shares (or even fewer) at a time. All this splitting and routing of a big order is complicated but easily automated. It is of course far cheaper to have computer systems do all of this than to pay human traders.
The goal of high-frequency trading (HFT) programs, running on computer servers inside data centres, is simply to trade profitably, ideally without accumulating too large – and therefore too risky – an inventory of the futures, shares, bonds or currencies being traded. The designers of HFT programs usually want them to end the trading day ‘flat’: with no inventory at all. There are, broadly, two ways for an HFT program to make a profit. The first is called ‘market-making’. If a market-making program is trading Apple shares, for example, it will continually post competitively priced bids to buy Apple shares and offers to sell them at a marginally higher price. The goal of market-making is to earn ‘the spread’, in other words the difference between those two prices – in Apple’s case, a few cents; in many other cases, a single cent – together with the small payments (around 0.3 cents per share traded) known as ‘rebates’ that exchanges make to those who post orders that other traders execute against. The purpose of these rebates is to encourage market-making, in the hope that competition among market-makers will encourage lots of keen pricing and thus attract other participants to the exchange.
The other way that an HFT program can seek to trade profitably is to be one of those other participants: it can ‘hit’ a bid or ‘lift’ an offer that is already in the order book. That’s often called trading ‘aggressively’. It’s more expensive than market-making: you have to pay the exchange a fee, rather than earning a rebate, and, if prices don’t move, your program can end up simply ‘paying the spread’ to market-making programs, because it will have to sell more cheaply than it buys. However, if an HFT program can identify a trading opportunity larger than the ‘spread’ (a high probability that, for example, the price of the shares being traded is going to rise or fall by several cents), then it may well need to act immediately and aggressively, before other programs do. There’s quite a bit of circulation of staff among HFT firms, and consequently the chief ways of identifying trading opportunities are common knowledge across the firms.
How is it that a computer program can identify a profitable trading opportunity? Both macroeconomic and company-specific news is often issued in machine-readable form, and a well-programmed machine can respond to it at least a quarter of a second faster than a human being can. There’s also increasing interest in using automated analysis of social media to gain a trading advantage. However, mainstream high-frequency trading programs rely primarily on two other sources of information.
The first is the content of the order book of the particular exchange in whose data centre the HFT program is running. As I’ve said, the exchange’s feed server continuously broadcasts changes in the order book (new orders, cancellations of orders, consummated trades). In US share trading, it isn’t the only way of keeping track of the market – you can also subscribe to the official, multi-exchange ‘consolidated tape’ – but the raw data feed direct from the feed server is faster, so any HFT program needs access to it and the program’s owner has to pay the exchange for that access. I’m told that the Merc charges $10,000 a month for access, and that this is cheap: many exchanges charge more. Monitoring the raw data feed involves keeping track of whether the order book contains more buy orders than sell orders, or vice versa, and whether the numbers of either are changing fast. Sometimes, too, a program may be able to detect a distinctive pattern in the arrival of new orders that might indicate that a big institutional investor is buying or selling a big block of shares.
In the early days of high-frequency trading, when execution algorithms handling institutional investors’ orders did dumb things like send in a child order every sixty seconds exactly, detecting such patterns was easy. Now, it’s more of a dark art. Who does it, with what success and how are very difficult questions to answer. Even those who do it may not fully know whether or not they’re doing it. In a field of complex electronic interactions, it may be hard to distinguish between a program that is successfully identifying and exploiting patterns of orders that result from the splitting of one big order, and a program whose success is based on less specific order-book patterns.
Unquestionably, though, the fundamental techniques of making predictions based on the balance between bids and offers in the order book are well known to all HFT firms. The resulting need for an advantage in speed is a major reason why HFT firms all have to pay for the fast, raw data feed. They have no choice other than to rent space for their computer servers in a data centre, because they need to be able to receive the data feed as quickly as possible and respond with orders – along with cancellations of orders that have been rendered ‘stale’ (mispriced) by changes in the market – that will arrive at the matching engine with minimum delay. Some data centres, such as the Merc’s, don’t allow you a spatial advantage within the data centre: if your cage is physically closer to the order gateways, the cable that connects your servers to them is coiled so that it is the same length as everyone else’s. In other data centres, I’m told, you can get closer by paying a higher rent. You also pay more if you want a ‘ten gig’ – ten gigabits, or ten billion binary digits, per second – connection to the order gateways and feed server; the standard connection is just one gig.
The second source of information used by an HFT program is order books other than that of the exchange in whose data centre it is running. Since the trading of shares in the US is spread across multiple venues, a change in one venue’s order book often presages changes in the others. That’s why HFT firms need the fastest possible communication links between data centres. If, for example, an HFT program is trading Apple shares in Nasdaq’s data centre in Carteret, it will most likely need access to a microwave link from Aurora. It will almost certainly also need access to the order books of the other exchanges on which Apple shares are traded. The firm that owns the HFT program doesn’t need to organise that access itself: Nasdaq, for example, will sell it to you. According to the most recent price list I’ve been able to find, you can expect a transmission time over the millimetre-wave link from BATS in Weehawken to Nasdaq in Carteret of around 105 microseconds, at the cost of $7,500 a month. The link from Direct Edge takes 101 microseconds and also costs $7,500 a month; a new link from the New York Stock Exchange is expected to take 190 microseconds and cost $10,000 a month.
Here, one has to look at maps to follow what’s going on. Michael Lewis, in his new book about high-frequency trading, Flash Boys, recounts what he was told by his main informant, Brad Katsuyama, who worked in the Royal Bank of Canada’s offices in New York. Katsuyama would, for example, often buy a large block of shares from one of the bank’s institutional investor customers, and then need to sell those shares in smaller blocks. He would frequently try more or less simultaneously to hit multiple bids or lift multiple offers for shares in the order books of several different exchanges. This would work for BATS – his orders there were typically executed in full – but by the time his orders reached other venues the bids he was trying to hit or offers he was trying to lift would have vanished from the order book. ‘The market as it appeared on his screens,’ Lewis says, ‘was an illusion.’
This is perhaps the most vehement complaint made against high-frequency trading: it leads to a vanishing market, one that disappears as soon as you attempt a trade. You can begin to understand why this happens if you take a taxi from Manhattan to Newark Airport. Quite likely you’ll be driven through the Lincoln Tunnel underneath the Hudson River. When you emerge in New Jersey, the first place you come to is Weehawken. The main fibre-optic cables from Manhattan to New Jersey follow the same route through the Lincoln Tunnel. The office in which Katsuyama worked was in Manhattan, like those of most of the big banks that operate as stockbrokers in the US. So Katsuyama’s orders were reaching the BATS matching engines at NJ2 in Weehawken before they reached the data centres of the other exchanges. Somehow, HFT systems trading in those data centres were learning of his other orders before they arrived, and buying, selling or adjusting their price quotes accordingly, causing the market Katsuyama could still see on his screen to have disappeared in reality.
I’d heard such complaints myself, and I’d been sceptical, because they seemed to violate the laws of physics. An HFT system in the Nasdaq data centre in Carteret, for example, can’t learn by magic about a child order that has arrived at BATS in Weehawken. There are essentially two ways it can find out: from the BATS raw data feed once it arrives in Carteret, or from a message from one of the same firm’s servers in Weehawken, which has learned that one of its bids or offers has been executed against. Even with a millimetre wave or laser link, all of that takes time: more time than it would take for the child order sent to Carteret to get there if it was travelling on the fastest fibre-optic route. I was put right by a physicist I met, who knows a lot about speeding up communications. My mistake was to have assumed that banks would know how to find the fastest cables, and would ensure that their own orders and orders from their institutional investor customers travelled down them.
Flash Boys has been much talked about, but no one seems to have focused on its crucial sentence, in which Lewis reports the difference in time taken between an order sent from the Royal Bank of Canada’s office in Manhattan to BATS in Weehawken and one sent to Nasdaq in Carteret: around two milliseconds (two thousandths of a second). That sounds tiny – no human being could perceive a time interval that short – but in the world of high-frequency trading it’s an eternity. Two milliseconds is ample time for news of the arrival of the order in Weehawken to reach HFT systems in Carteret: it’s about ten times longer than you should need if you’re using a fibre-optic cable that is optimised for speed and follows the most direct route.
Behind orders from banks’ institutional investor customers are people’s savings and pension funds. Flash Boys has been widely read as a morality play, a story of evil-doing high-frequency traders. But it can just as easily be read as an account of banks that either wouldn’t, or didn’t know how to, take best care of their own or their customers’ orders. To their credit, the Royal Bank of Canada team took action once they saw the disadvantage they were labouring under. I am assured by a source that other banks have done things to reduce the problem, for example moving their smart order routers from Manhattan into the data centres in New Jersey. All things considered, I suspect that what drains most money from pension funds and other savings are the high fees charged by those who manage them, and the excessive trading they often engage in, not high-frequency trading or even the incompetent handling of orders.
Is there an equivalent to the Lincoln Tunnel in Europe: a geographical quirk that creates exploitable predictability in the arrival of orders at different data centres? I don’t yet know: the geography of Europe’s automated trading is an issue I’ve never seen discussed in public. Two of Europe’s five main trading data centres are in London: the London Stock Exchange’s own centre and the Reuters foreign exchange trading centre, which is in Docklands. (I’m told that, unusually, trading firms originally couldn’t place their servers in the building containing the Reuters matching engines, which made locations close to that building very valuable real estate.) Two other data centres are just outside London. Equinix’s LD4, in Slough, hosts BATS Europe and one of the three global sets of matching engines of EBS, Reuters’s main rival as a foreign exchange trading venue (EBS’s other two sets are in NY4 and Tokyo). A data centre in Basildon, now owned by the US-based IntercontinentalExchange, contains the matching engines for Liffe (the London International Financial Futures Exchange) and for much of the trading of continental European shares. Europe’s fifth main financial data centre is Equinix’s FR2 in Frankfurt, which hosts the matching engines of Eurex (Europe’s leading futures exchange) and of the Deutsche Börse.
Europe’s data centres are connected not just by fibre-optic cables but also by microwave links, and those in and around London by millimetre wave links too. London’s rain is different from New Jersey’s – I’m told the average droplet size is smaller – making life in London somewhat easier for millimetre waves and, in the words of one of my contacts, ‘much worse for lasers’. So if you’re a Londoner, and are spooked by the idea of lasers flashing stock-market data overhead, be grateful for drizzle.
Go beyond the Chicago-New Jersey and London-Frankfurt clusters and you’re soon back in the slower world of fibre-optic cable. Oceans are a major barrier to microwave, millimetre wave and laser transmission through the atmosphere. Because of the curvature of the Earth and attenuation of the signals, all three require a series of ‘line of sight’ repeater stations. I’ve been warned not to imagine that this barrier can’t be overcome: serious attention is being given to such ideas as suspending microwave repeater stations from a series of balloons.
For now, though, transatlantic and other global financial trading links run through undersea cables. The routes followed by existing cables are often at some distance from the relevant geodesics, in order to minimise the extent to which they had to be laid on continental shelves. In shallow waters, cables have to be buried in the ocean floor, which is expensive, because they would otherwise be vulnerable to trawlers and ships’ anchors, and to attack by sharks (attracted by the electromagnetic radiation from fibre-optic cables). So an obvious thing to do is to lay new cables closer to the geodesic. It’s brutally expensive: laying a transatlantic cable can cost more than $300 million, and to my knowledge nobody has done it since the dotcom boom ended. A firm called Hibernia Networks planned to, with the intention of shaving around 2.6 milliseconds off the one-way transmission time on the fastest existing cable, Global Crossing’s AC1. But among Hibernia’s partners was the Chinese equipment maker Huawei, and the project seems to have hit trouble because of US cybersecurity concerns.
But you can do quite a lot without a new cable. You can add microwave links between the landing stations of existing undersea cables and the world’s main financial data centres. Doing that has reduced the total one-way transmission time from Aurora, Illinois to LD4 in Slough or FR2 in Frankfurt to 35-38 milliseconds. That gives HFT systems in those data centres usefully up-to-date information, about the order book for the E-Mini, for example (which is helpful not just to those trading US shares, but also to those trading European shares and futures).
As well as needing links of this kind, HFT firms have to speed up their computing as much as possible. Even the fastest conventional computer hardware is now not fast enough. At one HFT conference I attended, a firm was promoting a technology that involved submerging computer systems in liquid nitrogen, which permits ‘superclocking’: getting a computer system to run faster than would be safe at room temperature.
As far as I can tell, however, that approach hasn’t been much adopted by high-frequency traders: if nothing else, installing a liquid nitrogen system in your ‘cage’ in a data centre is likely greatly to increase the rent you have to pay. Much more widely used are specialised forms of computer hardware known as FPGAs, or Field Programmable Gate Arrays. The basic idea is to shift as much computing as possible off the microprocessor chips that make up the heart of a computer system, and to do that computing not by software but in fast FPGA hardware. I was told in Chicago in April that there was a buzz among the city’s high-frequency traders about the FPGA technology of a firm called Solarflare. This enables you to circumvent the kernel – the programs that manage the computer’s central processing unit, memory and input/output devices – and send bits (binary digits) directly from the raw data feed to designated locations in a computer’s memory. Solarflare ‘promise you a 1.2 microsecond one-way trip from the network to your user memory’, a programmer told me.
You can’t do everything using FPGAs: sometimes high-frequency trading requires use of a computer server’s central processing unit. The programmer tried to explain to me the software style that was necessary to ensure speed. ‘There are rules you need to follow to write fast code,’ he told me: ‘Don’t touch the kernel. Don’t touch main memory … Don’t branch.’ (That third commandment means don’t fill your program with ‘if’ statements – ones with the generic form ‘if A then do B else do C’ – because they get in the way of a modern microprocessor’s capacity to do several things in parallel.) But he also warned me that ‘we don’t know ahead of time what those rules are because every piece of code comes with a different rule book … I call them rules, but they’re more guidelines … That’s why it’s hard to teach someone: they either get it or they don’t. Those that get it, awesome.’
I’ve heard the requisite style of programming described as ‘bit fucking’. I’m uncomfortable with the term, but it conveys the need for an intimate understanding of precisely the best, subtlest way to handle the flow of binary digits from the raw data feed, through the computer system and the additional specialised hardware, and then back (in the form of orders) into the network connection to the relevant order gateway. A more polite term for it is ‘close-to-the-metal programming’. Just as talk of globalisation in finance tended to assume that local phenomena such as London rain had become unimportant, so it is all too easy to think that digital financial markets are abstract and virtual. But, like bankers who never look at maps, programmers who think that way will do a poor job in high-frequency trading: they will never ‘get it’. If you’re going to write really fast code, you have to understand the computer you are programming not as an abstract machine, but as a physical device through which electrical signals pass. Only then can you work out the most efficient way of channelling those signals.
If you’re a certain kind of person, there’s pleasure to be had in a lot of this. I don’t mean simply in bit fucking, although pleasure is of course one of the term’s connotations. There’s pleasure in skilled engineering, and in working for a firm with at most a few dozen employees that can outwit the big banks – which may be rich, but are also often bureaucratic and sometimes simply stupid. (Nearly all HFT firms started small, and only a few have grown to be more than medium-sized.) I confess that some of the pleasure rubs off on me. It’s nice to study a domain of economic life that’s so caught up with the physical world: with wind and rain and fog, tunnels and oceans and sharks; and with the geography of such unfashionable places as Aurora, Weehawken and Slough.
It should be said that there is less money to be made from high-frequency trading than you might think. If your program is market-making, for example, you might hope that it would make a couple of cents for every share it buys and sells; even a medium-size HFT firm can be trading a billion shares a month, so the cents would really add up. However, in this domain prediction is almost always probabilistic. For example, an HFT program seldom really knows that a particular little order is a child of a bigger parent: it can only guess. More generally, most programs’ predictions will be wrong almost as often as they are right. A market-making program is doing pretty well if it turns a profit of a tenth of a cent per share traded. There are, I’m told, opportunities that are quite a bit more profitable than that (and it’s these that ‘aggressive’ programs live off), but they are less common.
A good high-frequency trading firm can almost always make money, but costs of the kind I’ve listed weigh heavily, and you can’t avoid paying at least some of them at each of the many venues at which you are trading (the US has 12 share-trading exchanges and dozens of dark pools). The fact that a lot of these venues use the same handful of data centres – I’m told NJ2 in Weehawken is a favourite of dark-pool operators – creates some economies of scale, but each additional venue to which you feel compelled to connect adds costs. I get the impression that if an HFT firm fails it is most usually because it slowly drowns, submerged in a sea of costs.
Occasionally, an automated trading firm blows up spectacularly, rather than quietly drowning. The most dramatic case was the US market-making firm and stockbroker Knight Capital, which lost $440 million in 45 ghastly minutes on the morning of 1 August 2012. An old, long-unused trading program mistakenly left on one of its servers suddenly sprang to life, while the piece of program that kept track of the execution of its orders no longer worked. So the trading program kept on pumping more and more orders into the market. Knight’s staff wrongly guessed that the fault was in newly installed trading software, so they lost valuable time uninstalling it from their servers. By the time the old code was found and switched off, the firm was on the brink of bankruptcy.
Such events don’t always become public. In a New York coffeehouse, a former high-frequency trader told me matter of factly that one of his colleagues had once made the simplest of slip-ups in a program: what mathematicians call a ‘sign error’, interchanging a plus and a minus. When the program started to run it behaved rather like the Knight program, building bigger and bigger trading positions, in this case at an exponential rate: doubling them, then redoubling them, and so on. ‘It took him 52 seconds to realise what was happening, something was terribly wrong, and he pressed the red button,’ stopping the program. ‘By then we had lost $3 million.’ The trader’s manager calculated ‘that in another twenty seconds at the rate of the geometric progression,’ the trading firm would have been bankrupt, ‘and in another fifty or so seconds, our clearing broker’ – a major Wall Street investment bank – ‘would have been bankrupt, because of course if we’re bankrupt our clearing broker is responsible for our debts … it wouldn’t have been too many seconds after that the whole market would have gone.’
What is most telling about that story is that not long previously it couldn’t have happened. High-frequency firms are sharply aware of the risks of bugs in programs, and at one time my informant’s firm used an automated check that would have stopped the errant program well before its human user spotted that anything was wrong. However, the firm had been losing out in the speed race, so had launched what my informant called ‘a war on latency’, trying to remove all detectable sources of delay. Unfortunately, the risk check had been one of those sources.
Pleasure and risk are important, but we do need to come back to money. The right question to ask about high-frequency trading is not just whether high-frequency traders are good or bad, or whether they add liquidity to the markets or increase volatility in them, but whether the entire financial system of which they are part is doing what we want it to do. Of course, we want it to do several things, but I’d say that high on the list should be putting people’s savings to the socially most productive uses, while preventing too much of those savings being wasted along the way.
Speaking a couple of years ago to Bloomberg Businessweek about the new, faster cables, such as those planned by Hibernia, Manoj Narang, founder of the HFT firm Tradeworx, commented: ‘Nobody’s making extra money because of them: they’re a net expense … All they’ve done is impose a gigantic tax on the industry and catalyse a new arms race.’ The chief economic characteristic of an arms race is that all the participants have to spend more money, and none of them ends up any better off because of it. The programmer I spoke to in Chicago told me that his firm had to spend ‘godawful amounts of cash on IT’.
He had a simple proposal for how to stop the waste. The Securities and Exchange Commission (SEC), which regulates share trading in the US, should rule that matching engines couldn’t search for matches all the time, but only every hundred milliseconds. That would turn what is in effect a continuous auction conducted by the matching engines into a series of what economists call ‘batch auctions’. Some care would need to be taken over the way batch auctions were introduced. The feed servers would have to fall silent during the period between them, because otherwise the arms race would simply concentrate in its final millisecond. More generally, changes in a fast, interactive, highly complex system of the kind I’ve described can have unexpected side effects. But a regulator such as the SEC could introduce a change such as this on an experimental basis, for a limited number of stocks, and see whether it had benefits.
Variants of the batch auction proposal are being tried out in foreign exchange, but that’s a very different context. The big banks retain market power there that they have largely lost in share trading; the difficulties they have with HFT give them good reason to want to slow it down. The programmer wasn’t optimistic about the prospects of batch auctions in share trading. ‘It’s a dream, but it’s never going to happen’: too many people benefit from the current set-up.
I don’t think my programmer knew it, but his proposal for auctions every hundred milliseconds has an intriguing genealogy. A hundred milliseconds – a tenth of a second – is approximately the threshold of human perception of time, and so has played a significant role across a variety of human and physical sciences, not to mention cinematography.It’s a marker of how fast finance has become that being able to trade only every tenth of a second would now count as slow trading.