How Does High-Frequency Trading Make Money?

How does HFT make money?0:00

How do you make money from a stock without having any opinion about the stock? Say the only stock you have ever bought is an S&P 500 fund, a little every month. Then you hear about firms making millions trading the same stocks in microseconds, and it sounds like magic. Stocks are supposed to be owned. Not traded. Especially not if the ownership cycle is shorter than a blink of an eye.

The firms that do this are known as High Frequency Trading. Or HFT. So where does that money come from, how are they profiting, and is some of it yours? Are they genuinely building an infrastructure that is productive? Should we spend billions of dollars of capital to let a small number of firms pass around price quotes faster than a livestream shopping event? That is what we will find out.

Every other kind of trader has a view. The company is undervalued, the chart is about to break out, the internet is about to notice. High-frequency trading has no view. It trades the plumbing, like the queue of orders waiting at each price, the price difference between exchanges in milliseconds, and the fact that someone, somewhere, is about to trade, and you can be standing in front of them or out of their way.

I am now going to simulate that plumbing. Here is $STONK, and it will represent our fictional ticker on one exchange, at 1 millisecond resolution. You observe a market maker, the participant whose entire job is to be on both sides of the book at once. This is the expert screen. If you don't understand the charts, stay with the story and come back to this screen after you've heard it.

So this whole big screen presents only a single session of 20 seconds. 20 seconds of the book up close. Then, we also set up 3 different makers on the exact same order flow. The idea is to learn any differences in results by behavior. And the pro screen is the whole session the way a desk would look at it. So here, every row is a price, every column a moment.

The red and green brightness is how many shares were resting there, waiting. Most P&L charts show you what you made. This shows the thing you were standing in. Then, with every trade, we also add a nice colour to label who took it, noise, informed or sniper because the simulation assumes you are the admin. In real life, the flow is never labelled. You don't have nice colors telling you who bought or sold something.

Every trader you have heard of holds a view on value, and they differ mostly by how far ahead they look and what evidence they use. WallStreetBets asks, is this stock going to the moon. Their information is on vibes and momentum. They hold the stock whenever. But commonly you will hear them Hold On for Dear Life, which is another way of saying they never exit their trades and baghold anything they own until their grave.

A venture capitalist asks, what could this be worth. The information that interests them are TAM, CAC, LTV, Margins, and holds for years. An equity manager asks, what is this worth, and holds for months to years. A technical trader asks, where is the price going, from the price chart itself, and holds for days. A social-arb trader asks, where is attention going, from mentions, trends and TikTok, for hours to days.

Finally, the High Frequency Trader asks, who is about to trade, can I be in front of them or out of their way, Their information is the order book, the queue and latency, and they hold for microseconds to minutes. From the whole list, HFT is the only trader that trades flow instead of value. It makes the same money whether $STONK is a great company or a fraud, because it never holds $STONK long enough for that to matter.

WallStreetBets and HFT are the two ends of one axis. One holds forever on no information, the other holds for milliseconds on nothing but information. HFT names a speed tier, not a strategy. Inside it live at least 3 different kind of businesses, Market making, latency arbitrage, and very short horizon prediction Each of them make money in different ways.

How it happened3:22

Before we get into their mechanism and the money part, I think it is worth asking, how did high frequency trading happen? Sure, the money part is interesting, but why and how did people think it's worth doing? For that, we have to go back to 1971. NASDAQ started as an electronic quote board when humans still traded by phone. For the next 16 years, it looks like nothing interesting happened. But those were the years the plumbing got wired.

In 1975 Congress wrote it into law that every US stock market should be linked by computers into one national market, and that May the SEC ended the fixed commissions every broker had to charge. Over the next decade the wiring went in. A link between the exchanges in 1978, live trade reporting on Nasdaq and S&P 500 futures in Chicago in 1982, and in 1984, Nasdaq's Small Order Execution System, a machine that filled small orders at the dealers' posted prices without a phone call.

By the mid-1980s, computers were already trading the whole S&P 500 against its futures in Chicago, buying wherever it was cheaper and selling in the other market. Then on October 19, 1987, the Dow fell 22.6% in a day, and New York's electronic order line jammed on the paper printers at the trading posts. During the 1987 crash, many dealers stopped answering their phones, so in 1988 the Small Order Execution System was made mandatory for every dealer in Nasdaq's big stocks, so retail could always get filled.

It was then that a group of traders quickly figured a way to exploit it faster than the human dealers could update their quotes. The SOES bandits were among the first people to make a living from being quicker than the other side's screen. In other words, it was one of the first ways to make a living purely from being faster. High frequency trading, by hand. However, the screens were electronic but the trading was still manual. You had to sit in a day trading firm's room, at a Nasdaq terminal, with your eyes on the quotes and your fingers on the keys.

After 1988, a few small firms, many around Broad Street in lower Manhattan, let ordinary traders sit at those terminals and fire SOES orders the instant a dealer's quote fell behind the news, usually holding the shares for only a few minutes. The dealers called them bandits and handed out SOES Sucks buttons at their 1991 meeting, and the rules they pushed against them kept coming, until a federal appeals court sent the main one back in 1993, because the SEC had never measured whether the bandits really made spreads wider.

The dealers' lasting answer was software, programs that moved their quotes the moment a bandit hit them, and the easy trades dried up. Meanwhile, a programmer in one of those bandit firms, Datek, turned his order-watching program into Island in 1996. It was a fully electronic order book that paid rebates and let traders park their computers in its building, and a 2014 history says a rewritten version of its code still runs inside Nasdaq's matching engines.

The bandits got the blame for Nasdaq's wide spreads, but when a 1994 rule cut their trading, the spreads didn't narrow. That May, newspapers reported two finance professors' finding that dealers almost never quoted odd eighths in 70 of 100 big stocks, which kept the gap you paid at a quarter dollar or more. Within days, dealers in Microsoft, Apple and Cisco started quoting eighths, and their spreads fell by nearly half.

Investigators later found the dealers had been quoting better prices to each other on Instinet, a private screen the public couldn't see. The firms settled, agreeing to have their traders' calls taped for the Justice Department without admitting anything, and paying 1.027 billion dollars to investors in a class action. The SEC's fix, the 1997 order handling rules, forced those hidden prices into the public quote. Nasdaq's quoted spreads fell about 41%, and Island grew from about 3% of Nasdaq trades in 1997 to almost 13% by 1999.

Electronic communication networks took off from there, and the SEC wrote them a rulebook of their own in 1998. Two regulatory changes then built the modern market almost by accident. The first is decimalization which happened in 2001, The NYSE in January and NASDAQ in April, replaced fractions with pennies, and the smallest possible bid-ask spread, the floor under the human market maker's whole income, collapsed from a sixteenth of a dollar to a cent.

Overnight, your potential profit doing this market making business decreases by up to 6 times. The only way to make a living on a 1 cent spread is volume, and volume means machines.

Then Regulation NMS, adopted in 2005 and phased in through 2007, which banned any venue from trading at a worse price than the best one showing on another, so orders had to chase the best price wherever it was. Trading spread across a dozen-plus exchanges, and whoever could see all of them fastest got a permanent arbitrage. By the end of that decade, high-frequency traders were roughly half of US equity volume.

The SEC's own 2010 review said the estimates typically are 50% of total volume or higher. The rest is when the arms race starts becoming physical. When everyone starts being able to see prices from everywhere simultaneously, clicking fast enough doesn't cut it. Now, your order actually need to be faster than everyone else. For ordinary internet users, you click something and a page loads and get sent to you. No big deal.

It is a big deal for the high frequency traders. It is such a big deal that they think it's worth spending 300 million to buy a nearly straight fiber cable dug from Chicago to New Jersey in 2010. The latency of their trades after that investment went from 16 to 13 milliseconds. To understand this absurdity, 16 ms is a very good ping to be playing in League of Legends. You can barely notice a ping between 16 to 13.

Then, microwave towers, because fiber was too slow. Then a book, Flash Boys, in 2014, and an exchange built around a deliberate 350 microsecond speed bump, IEX, approved in 2016. Microseconds. That is about 0 ping in a League of Legends game. Along the way HFT firms simply became the market makers that exchange specialists used to be. Okay, we learned how these people got there. They found an arb, and spent billions racing for it. Whose lunch have they eaten?

For starters, it would be the day traders and scalpers. A day trader would look something like this girl from the FX Fighter Kurumi-chan anime. She would be on coffee all day, staring at the screen, trying to time her fills and praying to god while the candle jitters. Their edge was the spread itself, buy the bid, sell the ask, pocket a sixteenth, until the sixteenth became a penny, and the penny went to the machines.

Even if the manga had been written 20 years earlier, I don't think Kurumi-chan would have stood a chance against the machines, because by 2010 the SEC said estimates put HFT at 50% of US stock volume or higher. The next people HFT killed is the exchange specialists and dealers who had a human monopoly on the order flow at their post, It wasn't until decimalization and then Reg NMS that gave the post to electronic makers.

The SOES bandits were quicker than a dealer's hands, until machines were quicker than any hands. Banks' arbitrage desks spotted the same thing priced differently in two places, and the same job got automated. All of these competition and race to the bottom yielded an unexpected winner. The long term investors, whose edge is patience, came out as net winners. They get the lowest transaction cost in history.

The order book10:05

Finally, we get to the part of its engineering. How does HFT work? For that question, we must understand the stock exchange. For each price, the exchange keeps a queue of orders waiting to trade there, Bids below the current price, and asks above it, On most US stock exchanges, the transactions are first come, first served. This rule is very important for because it is the entire premise of latency arbitrage.

A limit order joins the back of the queue at its price. A market order sweeps the best queue first, then the next, until it is filled. Whoever is at the front of a queue when a market order arrives gets the trade. This is price-time priority. So, queue position is an asset. Two orders at the same price are not equal. The one at the front fills on the very next order that comes through, small or large.

The one at the back fills only when everyone ahead of it has been taken or has cancelled, usually by a big order, And big orders are the ones most likely to be informed. Being last in line means you get filled precisely when you would rather not be. The spread is not a fee, it's a price. The gap between the best bid and the best ask is what the market pays a maker to stand there.

It's set by competition among makers, and it goes as low as adverse selection, the risk of getting sniped, inventory risk and running costs allow, Or until it hits the 1 cent tick, That is where makers stop competing on price and start competing on queue position. Let's watch it happen. Here is $STONK, and here is our maker, offering to buy just under the price and to sell just over it.

Every setting in this toy market, the prices, the sizes, the speeds, is our own choice, so don't read the numbers, read what happens. A buyer with no special information comes in and sweeps the asks, and our ask is there, so we sell. But we weren't first in line, and other orders at our price were filled before ours. Over the whole session our maker trades again and again, almost always with somebody ahead of it in the queue.

Now look at our worst fill of the session. It stood at the very back of a long line, so it only filled when a big order ate the whole line, and a second later the price had moved against us.

Three ways to earn11:55

Now, how does HFT make money? There are 3 ways, ordered by impact. The first is market making, selling the spread many thousands of times. Quote a bid and an ask around the mid. When a buyer arrives you sell at the ask, when a seller arrives you buy at the bid, That difference is yours. For $STONK, that difference is a cent. When you multiply by over a thousand fills per stock per day, across thousands of stocks, and the cents add up to become a business.

Then the exchange adds a rebate for providing the liquidity that got taken. This rebate mechanism is especially important, because it is another source of income from market making. In our simulation it makes up a large chunk of the profit. In real markets it's messier. One study of Nasdaq found that resting orders only made money once the rebate was added. Others find the rebate mostly gets competed away into tighter spreads.

For more details, check out this reading list. A 0.2 cent rebate for the maker and a 0.3 cent fee for the taker, per share, are the US convention. What this means is if you wait in line, you are the liquidity "maker" and get paid a "free" 0.2 cent by the exchange. If you want your trade right now, you are a liquidity "taker" and you pay a 0.3 cent tax. And the fastest firms happily pay that tax, whenever a stale quote is worth more than 0.3 cent.

So what makes you a maker or a taker? A maker posts an order and leaves it waiting in the queue, so it adds something to the book that others can trade against. A taker trades against an order that is already waiting, so it takes that order away. The catch is that a maker's quote is a promise standing in public. When the price moves, that promise goes stale, and the maker has to cancel it before somebody takes it.

That race between your cancel and someone else's order is the cancelling game, and we'll watch it in a minute. And if you're slow, your new quote can even land on the wrong side of the price and trade on arrival, so you become a taker by accident and pay the fee instead of earning the rebate. In the meantime, market makers watch out for adverse selection. We measure it as how far the mid price moved against us between the moment of the fill and 1 second later.

When an uninformed trader hits you, the price wanders and you keep the spread. Low risk money. When an informed trader hits you, someone who knows where the price is going, the price keeps going through your quote, and the fill was not income. It was the bad news arriving. A fill is evidence that you might have been wrong, and the market maker's real skill is telling the two kinds of fills apart and pricing the second into the spread.

By predicting the flow and choosing to either get ahead of it or dodge it. This is also why the big wholesalers, off-exchange market makers built on HFT technology, pay retail brokers for their customers' orders This practice is called payment for order flow, and it is a big part of how Robinhood can charge its users 0 commissions for trading. When it went public in 2021, Robinhood said these payments, plus its crypto rebates, were 75% of its 2020 revenue.

It doesn't sell information about your order. It sends the order itself to a market maker like Citadel Securities, and the market maker pays Robinhood for it. In 2025, Citadel Securities alone was 13% of Robinhood's revenue. Why would a market maker pay for Robinhood's orders? Because retail flow is, on average, far less informed, so filling it is low risk. How do you tell informed flow from noise? You don't, not at the moment of the fill.

Going back to our simulation, we label every trade with nice colors because we are in god mode. In real life, a maker has to count thousands of fills and sort them by everything it can observe, which broker the order came through, its size, how soon after a news print it arrived, from which exchange, Then a toxic flow gets labelled after the maker recognizes the fills that kept losing.

Our simulation can cheat, so let's use it. Take the fast maker that doesn't defend itself, and sort its fills by who hit it. Against noise traders it earns a little on almost every fill. Against the informed trader it loses a little more than that, on almost every fill. And one single fill from the sniper costs as much as a whole stack of good fills put together. The second way is latency arbitrage, that is being first when the world moves.

Example. The same stock trades on many exchanges. An ETF and its basket are almost the same thing. A future and the spot it settles to are almost the same thing. When one of them moves, the others are momentarily wrong, and whoever reaches them first collects the difference. In the simulation, anyone who moves this fast, not including you, is the sniper. When news moves $STONK's true value, every quote left on the wrong side is stale, and the sniper takes all of them, at the old prices, before their owners can cancel.

Whether your quote is among them is decided by one comparison, Your latency against theirs. Who is this sniper? The sniper is an ultra-fast automated algorithm located directly inside the exchange's data center. It sits idle until a public news event occurs. The instant news breaks, it sweeps up the stale quotes still sitting at the old price. If your algorithm can't cancel outdated quotes faster than the sniper can execute against them, the sniper will take your money every time news hits.

Now watch what happens when news hits. The true value jumps, and every quote still sitting at the old price is suddenly a bargain. The sniper sweeps them a heartbeat later, and most of the shares that were waiting near the price, the liquidity, are simply gone. Our fast but naive maker had a quote in there, and it got taken, because being fast doesn't help if you never pull your quote.

The careful maker is just as fast, but it steps out the moment it sees news, so the sniper finds nothing of ours to take. The third way is short term prediction. The queue itself leaks information. Suppose the best bid has 10 times the size of the best ask. It's normal to assume the next tick is more likely to be up. Strategies that trade those milliseconds-ahead statistics are like trying to catch a wave before it arrives.

If you are a day trader, then this was your lunch, now abstracted by a script. Example, find a big candle, get filled to the next gap or something. Disclaimer, this is not financial advice, and strategies are never static. I am just describing that high frequency trading sometimes does things that are similar to day trader's judgment

The steamroller17:29

Now that we settled on its profit opportunity, it is worth asking. Isn't this pennies in front of a steamroller? By definition, yes. Tiny gains, and maybe a large loss. So how does the market maker not get blown up? For that question, we have to inspect those who blown up. The LTCM hedge fund that blew up in 1998, and lately our OpenAI alumni Leopold Aschenbrenner, whose leveraged AI fund lost two thirds of its value in July 2026, when the margin calls came.

They share one property HFT does not have Commitment. They hold a position they cannot get out of when the steamroller arrives. And yes, the crypto genius SBF blew up in 2022 too, but that wasn't a position he couldn't sell. His trading firm had been spending FTX customers' money, and a jury convicted him of fraud for it. The market maker is a penny-picker with no commitment, and the business is engineered around that fact.

First, risk is position times time, and they hold neither. A crash takes seconds to minutes to develop. A maker's exposure per fill is small and brief, and it doesn't keep what it buys. They also have a hard cap on their inventory. When it's long, its quotes shift so the next customer is more likely to buy from it than sell to it, so the position drains away on its own and they try to balance and end the trading day with 0 positions.

Second, the spread is the steamroller premium. High frequency traders are literally paid to stand in front of a steamroller and get out of the way before it hits. If no one dares to stand, the spread gets wider until it starts to be worth it. Third, they have the option to not be there. When news hits, or volatility jumps, a maker can quote more carefully, wider, smaller, or not at all.

If enough makers do that, the spread widens until some other maker is greedy enough to take the risk at the new price. In the 2010 flash crash, one huge automated sell order hit a thin market and many makers stepped back within minutes of each other. In some stocks nobody was left to take the other side, and prices fell through the hole to placeholder quotes, some at a penny.

Their safety is the market's fragility. Inversely, a functioning market requires a supply of daredevils or very smart risk takers to take any trade. Fourth, speed reassigns the steamroller to someone slower. When news hits, the fastest firm can cancel its quotes in a millisecond, and the same event then runs over whoever was still quoting. If you are fast enough, you can play hot potato with anyone without burning your own hand.

So the faster you trade, the lower the risk. Fifth, the law of large numbers reduces their risk. Trade hundreds of thousands of times, and your P&L settles into a range you can accept. Virtu's boss said only about half of its trades made money, and yet the firm reported 1 losing day in 1,238. So how does a high frequency trader blow up or lose money? Often, it is operational, not directional.

Example, A deployment bug sent millions of orders in 45 minutes and cost one firm about $440 million in 2012. Bad data feeds, exchange outages, or a risk module that fails at launch. The classic penny-picker is a man who picks up slowly from the road. The market maker is someone who runs fast enough to pick up any penny, and the only thing that reliably kills him is his own shoelaces.

The verdict20:24

Let us return to our simulation. Now, we inspect the rewards of being fast, and how fast we need to be to make certain decisions worth it. Same 20 seconds, same orders from everyone else, 3 makers. A human scalper at 100 milliseconds. This will be our control. A fast, naive maker at 1 millisecond. And a fast, careful maker at 1 millisecond that leans its quotes against what it holds and steps out for 150 milliseconds after news.

Now run it 40 times, in 40 different worlds. The careful maker made money in 35 of the 40 worlds. The scalper lost in 21 of them, and the naive maker, just as fast as the careful one, lost in 28. Here's our explanation The human scalper is uncompetitive. He is too slow to react. Its outdated quotes keeps getting picked off by faster participants, and when it re-quotes late its orders cross the book by accident, so it pays taker fees. It performs poorly overall, especially during news events.

The fast, naive maker is the worst performer. Because it is fast, it's always at the best price and gets filled constantly. But without safety rules, it gets repeatedly hit by informed traders right before the price moves against it, and that's adverse selection. The fast, careful maker is the consistent winner. It succeeds by using 2 defensive rules. Inventory skew pulls back its bids or offers when it holds too many shares, so it avoids over-exposure.

And the news pause steps it out of the market for 150 milliseconds during news events, to dodge sudden price shocks. The careful maker also held a quarter as many shares as the naive one, on average, and its worst losing streak was less than a third as deep. The moral of the story? Speed alone does not guarantee profit in market making. In our simulation, raw speed without rules simply increases your exposure to risk, while risk discipline protects your capital.

Real firms need both, and in the real data the faster firm earns more, partly because speed is what lets it step out in time. And look at what the careful maker does to everyone else. After each news jump, the shares waiting near the price fall to about half. The careful maker is part of that hole.

The speed race22:16

Latency is the time from something happening in the world to your order arriving at the matching engine. It's made of several parts, the feed reaching you, your decision, and your order. It matters because after news, every quote resting at the old price is free money for whoever reaches it first. Your quote survives only if your cancel lands before the fastest predator's order. Now, let us check out how much it's worth upgrading your latency.

The cost of being slow has a shape. Loss to snipers is roughly the number of news events, times how wrong a stale quote is minus the half spread it was paid, times the chance that your cancel arrives after the predator's order. Only the last factor depends on speed, and it's a cliff, not a slope. Faster than the sniper, and we are that sniper. Slower, and we are sniped at every news event.

The cliff decides whether we pay the news tax, not whether we can still be profitable. A sniped fill costs us 1 lot times the jump, minus the half spread we were paid, And whether that tax exceeds our income depends on How often news comes how big the jumps are, against how often we get filled and at what spread. Since ranking, not speed, determines the winner, the race has no finish line.

Every dollar spent trying to win the latency game is a deadweight tax on everyone who trades. The formal version of this complaint is Budish, Cramton and Shim in 2015, and the races have since been measured on exchange message data by Aquilina, Budish and O'Neill in 2022. Here is that cliff in our simulation. Each point is our maker's average profit per session at one latency, against a sniper with a fixed speed.

If you're faster than him, he never catches your quotes, so you lose nothing to him. If you're slower, you lose your quotes to him almost every time. Past the step you're already getting caught almost every time. So why does revenue keep going down as you get slower? Because slow quotes are wrong quotes, the price moved and yours didn't. And why does it flatten out instead of going to minus infinity?

Because a really slow quoter barely trades at all. Fewer trades means less to lose. You're not winning, you're just not playing. The line for the 20 millisecond sniper starts to fall way before 20 milliseconds. This is because the sniper isn't the only one hunting you. The informed trader may hit stale quotes too, and it reacts a few milliseconds after news at random times, not on a fixed clock.

With a slow sniper, the informed trader becomes the fastest hunter. A naive maker never generates revenue at any speed because it isn't cautious about informed flow. Its worst spot is 5 milliseconds, still fast enough to get filled constantly, too slow to be right about the price. Past 5 milliseconds its line recovers, not because it starts to make money, but because it quits. Both makers meet when their latency gets slow enough, because at 100 milliseconds everybody is the human scalper.

How much does it cost to be fast? In our simulation, we get to pick our latency with a slider. In real life, the slider is expensive. This is how High frequency trading became a high capex business instead of a contest of being smarter than your rivals, and why individuals should give up day trading. Even so, for the big firms the biggest bill isn't the hardware, it's the fees on every trade.

These are estimated numbers, not measurements. At 100 milliseconds, it's a person, a screen and a mouse. That's you, or any retail app, and it costs a laptop. At 20 milliseconds, it's a program on a rented server in another city, talking to the exchange over the internet. That's a hobbyist's algorithm, for hundreds of dollars a month. At 5 milliseconds, a server in a data center near the exchange, with ordinary software and network cards.

That's a small trading firm, for tens of thousands a year. At 1 millisecond, co-location. Your server sits in the exchange's own building, with the cable runs equalized so nobody's cabinet is a nanosecond closer to the matching engine, and the exchange's direct data feeds, which arrive before the public tape. That's a mid-size HFT, at hundreds of thousands a year in cabinets and feeds, before people. At a tenth of a millisecond, you need a kernel bypass.

The network card hands packets straight to your program with no operating system in between, the code is lock-free, and clocks are synchronized to nanoseconds so you can even tell who was first. That's most HFT firms, and the cost is engineers. At 1 microsecond, you are looking into FPGAs. Chips wired to read the feed and fire the order without a CPU in the loop, about a microsecond from tick to trade, against tens of microseconds for well-written software.

You need to be the top handful of firms, with millions in hardware and more in people. And at the late game, you want quotes between cities using microwave. You want radio towers on the straight line between two venues, because radio through air is faster than light through glass. This is the infrastructure afforded only by a top few, costs hundreds of millions dollars

The limits26:33

Suppose after watching the whole thing, you want to build your own high frequency trading firm, there are something I should warn you about. Our simulation ignores too much of reality. First, our simulation uses only one venue. In practice, latency arbitrage lives between venues, where a price prints on one exchange and the race is to all the others. Our sniper here is a stand-in for that whole cross-venue race and they have a constant speed. In real life, they are scattered around with different reaction time, so you can't always be clear who is the sniper and where he is coming from.

Next, our sniper and the informed trader see the true value directly. In real life, predators infer it from correlated instruments and the queue. Our losses to them are higher here than in real life, because we let our simulated predator cheat. Third, our order flow is synthetic, so we have noise traders at random depths, limited informed trader, and each news triggers a price update. A real book has hidden orders and icebergs, thousands of participants with different interest, and trades that arrive in clusters.

Fourth, we haven't modeled tick-size rules, opening or closing auctions, halts and short-sale limits. In the real world, every one of these changes requires a maker to adjust their quote around it. Finally, every number in our simulation is invented. Fees, rebates, latencies and the size of a news jump are all approximations. The real numbers are proprietary tick data I'd get sued for sharing, so read the shape of the graph, not the numbers.

Now come back to the expert screen, and read its P&L panel line by line. Spread captured is the profit earned by buying slightly below the mid price and selling slightly above it. This is the money the job is supposed to make. Adverse selection asks, 1 second after each of our fills, where was the price? If it had moved against us, that gap times the shares is what smarter, faster traders took from us.

Maker rebates are what the exchange pays us every time our resting order gets hit, In our simulation, this is the baseline income that keeps the careful maker profitable. In real markets, a rebate can be the margin on a resting order, but for a big firm like Virtu, the exchange bill after rebates is still its biggest expense. It comes down to how much liquidity you want to draw or give, and how profitable you can do it.

Taker fees are what the exchange charges us when we hit someone else's order. In our simulation, paying them means our orders were too slow. Real fast firms pay them on purpose, whenever taking a stale quote is worth more than the fee. Inventory marked to mid is the value change of the shares we still hold at the end, and it reveals whether our P&L came from a real trading edge or simple price luck.

Session P&L is the sum of all of those. Average shares ahead in the queue tells you how far back we stood, and sitting far back means your fills are more likely to be during large, toxic price moves. Average wait to fill is how long our orders sat there, and long waits leave stale quotes exposed. Fills taken by the sniper are our direct cost of failing to cancel stale orders in time.

Time withdrawn is the share of the session spent with no quotes at all, the trade-off between dodging a crash and making an income. Worst drawdown is the largest drop from a peak, the number a risk desk uses to set capital limits. And liquidity left after news is how much depth survives a sweep, and when it drops to zero, flash crashes happen. So these are simulated numbers, but I'm sharing the kind of indicators a HFT would want to care about because that informs your performance at the end of day.

Your money29:34

So, back to the S&P 500 fund you top up every month. Is some of that money yours? Say you put $500 into an S&P 500 ETF like VOO every month. When you buy, you pay the ask, which sits about half a cent above the middle price, and on a share that costs around $700, half a cent is tiny. Each $500 purchase pays about a third of a cent in spread, so a whole year of buying costs you about 4 cents.

Now compare that with the fund's own fee. VOO charges 0.03% a year, which on your first year's balance comes to about 98 cents. And around decimalization, a big stock's quoted spread went from about 11 cents to about 6, and kept falling as the machines took over. So what about the race itself? Aquilina, Budish and O'Neill measured it on the London Stock Exchange in 2015. In the biggest stocks there was a race about once a minute, usually won by 5 to 10 microseconds.

Added up, the races cost about 0.42 basis points of everything traded, 4 thousandths of a percent, and across the world's stock markets they put the sums at stake at around $5 billion a year, most of it won by a handful of firms. Apply that rate to your $6,000 of purchases in a year, and it comes to about 25 cents. In the authors' own words, the tax "does seem small enough that ordinary households need not worry about it".

That doesn't mean nobody got hurt. On May 6, 2010, a mutual fund complex started selling about $4.1 billion of stock index futures. High-frequency traders passed the contracts back and forth like a hot potato, and the major indices, already down over 4%, fell another 5 to 6% in minutes. Over 20,000 trades went off at prices more than 60% away from where they had been moments before, some at a penny, and many of them were retail stop-loss and market orders.

The exchanges cancelled them. On August 24, 2015, IVV, one of the big S&P 500 ETFs, fell more than 20% below the previous day's close in the first minutes of trading. And the measured costs of speed land mostly on big orders. When high-frequency traders detect a large institution working an order, they can end up trading in the same direction, which makes that institution pay more. So, is some of that money yours?

Cents a year of it, at most. The fee on your fund is bigger than everything the race takes from your buys, and choosing between two S&P 500 funds can move more of your money than the whole race does. The real costs land on the big traders, and on anyone who sells with a market order in the middle of a crash. So how do you make money from a stock without having any opinion about the stock?

You stand in line on both sides, you step out of the way when the steamroller comes, and an order like yours, small and uninformed, is the one they're happiest to fill. And when someone else is too slow to step out of the way, the fastest of them take that too.

Sources

1. Brogaard, Hendershott & Riordan (2014). "High-Frequency Trading and Price Discovery." Review of Financial Studies 27(8): 2267–2306On Nasdaq in 2008–09, HFTs' resting orders lost money before the exchange's rebate and made money after it.
2. Menkveld (2013). "High Frequency Trading and the New Market Makers." Journal of Financial Markets 16(4): 712–740One big HFT in Dutch stocks, trade by trade: four in five of its trades were resting orders, and fees were a big part of its profit.
3. Malinova & Park (2015). "Subsidizing Liquidity: The Impact of Make/Take Fees on Market Quality." Journal of Finance 70(2): 509–536When Toronto changed who pays the fee and who gets the rebate, quotes moved and what traders actually paid stayed the same.
4. CFTC and SEC staffs (2010). "Findings Regarding the Market Events of May 6, 2010."The regulators' account of the Flash Crash, which began with one 75,000-contract (about $4.1 billion) E-Mini sell program.
5. Virtu Financial (2014). Form S-1 registration statementA real HFT firm's own numbers: 1 losing day in 1,238, and exchange fees after rebates as its biggest expense.
6. Aquilina, Budish & O'Neill (2022). "Quantifying the High-Frequency Trading 'Arms Race'." Quarterly Journal of Economics 137(1): 493–564London Stock Exchange messages: a speed race about once a minute per FTSE 100 stock, mostly won by the same six firms.
7. Baron, Brogaard, Hagströmer & Kirilenko (2019). "Risk and Return in High-Frequency Trading." Journal of Financial and Quantitative Analysis 54(3): 993–1024Swedish HFT firms, 2010–2014: what pays is being faster than rivals, not being fast, and the profits stay with a few firms.
8. Budish, Cramton & Shim (2015). "The High-Frequency Trading Arms Race: Frequent Batch Auctions as a Market Design Response." Quarterly Journal of Economics 130(4): 1547–1621Why a market that runs non-stop rewards tiny speed edges, and a proposed fix: auctions held many times a second.
9. U.S. SEC (2010). "Concept Release on Equity Market Structure." Release 34-61358The SEC's own description of HFT and its strategies: market making, arbitrage, structural and directional.
10. SEC Division of Trading and Markets staff (2014). "Equity Market Structure Literature Review, Part II: High Frequency Trading."A plain-language survey of what the HFT studies found, including that HFT is many strategies, not one.