High-Frequency Trading Is Good

Sichuan Mala has written a guest post on one of the most unfairly maligned parts of the financial industry

High-frequency trading (HFT) is one of the most unfairly maligned industries of the modern era. Despite being a source of market liquidity, price discovery, and technological innovation, HFT is often decried as a “waste of talent” which entraps the best and brightest in endless spirals of negative-sum competition. Even HFT practitioners themselves sometimes become disillusioned with their careers. It’s not difficult to see why: intuitively, it’s easy to buy the argument that quant-versus-quant competition down to the sub-microsecond level generates little value to anyone who doesn’t come out on top. However, a thoughtful examination of the costs and benefits of HFT suggests that it’s actually an incredibly pro-social industry: the world at large profits dramatically from the efforts of this relatively small industry. Let’s dive in.

---

What is HFT?

To begin with, it’s important to orient our understanding around what exactly HFT is and what high-frequency traders do. Generally speaking, the HFT industry can be separated into two major classes of trading firms, namely makers and takers:

• Makers (short for “market makers”) are market participants who quote both sides of a market, adding liquidity by placing both bids (buy X of asset Y at price Z) and asks (sell X of asset Y at price Z) into the order book.

• Takers, as the name implies, execute orders in the market, filling a participant’s preexisting bid or ask.

Making and taking are not inherently high-frequency; however, in the last 15 years, many makers and takers alike have found that being able to place or execute trades on short timescales is a significant market advantage. Competition occurs in all directions: makers compete with other makers to offer tighter bid-ask spreads; takers compete with other takers to capitalize on proprietary trading signals (alphas) faster; finally, makers try to update their prices fast enough so that takers can’t execute against stale bids or asks.

---

HFT Increases Liquidity and Reduces Transaction Fees

The simplest and most commonly cited benefit of HFT is the improvement to market liquidity and the reduction of bid-ask spreads, which lowers the costs paid by all market participants. Recall that a market maker profits by quoting a double-sided market (e.g., buying apples for $0.99 and selling apples for $1.01). Competition between market makers naturally drives down the bid-ask spread. However, holding the trading latency fixed, there is a natural limit to how far spreads can be reduced. This is because market makers are inherently exposed to risk in the form of fluctuations in the true price of the asset.

Imagine, for example, that you are quoting a spread of ($0.99, $1.01) in the apple market, and that you can only adjust your quotes once per hour, exactly at the top of the hour. Now imagine that news comes out at 12:30 PM that cicadas have devastated apple orchards across North America, driving the true price of apples to $2. Your quotes are now stale, because you’re quoting around a price which is no longer accurate, and because you can’t adjust your quotes until 1:00 PM, savvy traders will pick off your stale $1.01 asks and easily make a ~$1 profit. The more risk you take on, the more you will charge to make up for that risk; conversely, if market makers are able to adjust their quotes rapidly to account for new information, they will take on less risk and consequently feel comfortable quoting a tighter spread, allowing them to capture a greater amount of trading flow. With lower latency, makers will also feel more comfortable offering a larger amount of bids and asks, deepening the order-book liquidity and substantially reducing the cost of large trades.

This might seem like a contrived and academic argument. However, real-world evidence suggests that the introduction of HFT does actually lower transaction costs:

• Gerig (2012): HFT on the NASDAQ synchronizes prices of related securities, improving price accuracy and reducing transaction costs.

• Brogaard, Gariott, and Pomeranets (2018): The introduction of HFT in the Canadian stock market between 2008 and 2012 reduced spreads by approximately 0.8 basis points on average.

• Citadel Market Lens, March 2022: When the Taiwanese stock market transitioned from frequent batch auctions to continuous trading, lowered bid-ask spreads and deeper liquidity were estimated to save investors >$1.4 billion TWD per year.

While in absolute terms the reduction in the bid-ask spread appears minimal, it is important to remember that the bid-ask spread represents a cost paid by all market participants all the time they make a trade, and that large bid-ask spreads are especially punishing toward smaller investors, for whom transaction costs represent a greater proportion of trade size.

There are also compelling theoretical arguments why reducing latency, even well below the limit of human perceptibility, should allow market makers to reduce spreads:

• Market making is analogous to selling options: In a sense, submitting a bid or an ask to the order book is like writing a put or call option that can be executed against by the market, where (roughly speaking) the latency that the market maker faces is analogous to the time-to-expiry of the option. Standard Black-Scholes option pricing suggests that reducing the time-to-expiry should also reduce the premium that the option writer charges, in this setting analogous to the distance of the bid or the ask to the midpoint price of the asset.

• Reality is fundamentally continuous: Events in real life, which affect asset prices, happen instantaneously, rather than being conveniently binned into 1-millisecond or 5-second windows. Because market volatility often occurs in rapid response to real-world events, volatility is typically very “bursty” and tends to co-occur on even single-digit microsecond levels, suggesting that bid-ask spreads should be sensitive to latency improvements down to sub-microsecond levels.

Accordingly, we should expect bid-ask spreads to continue to narrow as market participants continue to engage in latency competitions. To be sure, there is some level of latency competition which is unlikely to result in any benefit to transaction fees. As a contrived example, it seems doubtful that spreads would be much different if latencies were 10 femtoseconds above the physical limit compared to if they were 5 femtoseconds above. But the fact that this is a contrived example in and of itself reveals the absurdity of this concern: due to the extreme capital costs and engineering challenges of pushing the low-latency technological frontier, it is unlikely that more than a small handful of firms would even bother making such large capital expenditures, and the diminishing returns on marginal investment serve to naturally limit the extent of such competition.

For example, around 2007, a large number of trading firms, such as Getco, invested in building fiber-optic lines between Chicago and New York to compete on arbitraging ES and SPY. This is the prototypical example of “race-to-the-bottom” dynamics that inevitably comes to mind when people think about HFT, largely due to the influence of relatively inaccurate outsider portrayals like Flash Boys. In reality, however, many of these firms overestimated the value of competing in this way, resulting in major cost overruns and an eventual scaling back of such expenditures. It is precisely because competing on such frontiers is so expensive and challenging, and the margins involved so small, that such competition is inherently self-limiting.

---

Latency Competition is Natural and Inevitable

A skeptic of the above argument might well ask: Isn’t this a circular argument, in that it presumes the existence of other market participants operating at faster timescales who could pick off stale maker quotes? Intelligent human decision-making operates at the level of seconds and higher; accordingly, if we all agreed to trade at, say, 100 millisecond latency at best, wouldn’t we all save a lot of effort and money, and not be noticeably worse off at all?

This is essentially the motivating argument behind the idea of frequent batch auctions (FBAs), proposed in 2015 by Budish et al. as a solution to the “HFT arms race.” The idea behind FBAs is that instead of continuous-time trading, where market orders are processed in the order received, FBAs “discretize” time into fixed blocks of, say, 100 milliseconds, and all orders that arrive within each block participate equally in that block’s uniform-price auction. While superficially compelling, FBAs are a deeply flawed solution for both practical and theoretical reasons, and a close examination of these reasons illustrates the inevitability of continuous-time latency competitions.

To begin with, the entire publication from Budish et al. is based upon a false premise. The authors identify that on the microsecond level, prices of the ES (E-mini S&P 500 futures) tend to lead changes in the price of the SPY (the New York S&P 500 SPDR ETF), and claim that when the price of the SPY moves, that typically represents a “latency arbitrage” where a low-latency taker snipes out-of-date bids or asks, moving the price of the SPY in line with the ES.

250 millisecond-level price evolution of the ES and SPY

In the world of FBAs envisioned by Budish et al., traders on the ES observe public information about the true value of the S&P 500 before traders on the SPY, but because both instruments are traded on coinciding batch auctions, there would be no divergence between the ES and the SPY and hence no opportunity for “latency arbitrage.” Essentially, between the emergence of new price information and the end of the current auction block, everyone could simply politely shift their bid-ask spreads to be centered around the new price.

One critical flaw in this argument is that price movements do not necessarily reflect the incorporation of external public information. For example, suppose a long-term trader has a large short position on the ES and wishes to close their position. The trader will buy E-minis, moving the ES price higher and potentially creating an “arbitrage opportunity” against the SPY, which is then closed by arbitrageurs. In this example, notice that the trade which induces price movement of the ES is itself the sole source of new price information: therefore, whether the market is continuous or batched, there will always necessarily be some kind of delay between the movement of the ES and the SPY, because there is no reasonable way to anticipate this price movement aside from direct observation of the movement itself! (In fact, it’s relatively likely that the trader bought E-minis from the same arbitrageur who moved the price of the SPY up: a beautiful example of how HFT “knits together” separate markets, improving both liquidity and pricing) Accordingly, the “costs” calculated by Budish et al. are, at best, highly suspect; at worst, completely illusory and based upon fundamental misunderstandings of the financial markets.

Moving beyond purely theoretical criticisms, the most compelling evidence against FBAs comes again from the Citadel Market Lens, March 2022, reporting on the effects of the Taiwan Stock Exchange (TWSE) switching from a 5-second batch auction format to a continuous time, price-time priority limit order book. The effects were dramatically positive:

• Bid-ask spreads were significantly reduced.

• The depth of the order book more than doubled, with 5-fold increases in liquidity depth for the most traded names.

• Finally, overall volatility dropped by 18%.

It’s unsurprising that usage of batch auctions decreases the amount of available liquidity and increases spreads. In an auction format, market participants have significant uncertainty about the final execution price of their trades as well as the current structure of the market; in return for taking on greater risk, makers must logically charge a higher fee for their liquidity. Moreover, in more complex instruments, market participants will often partially hedge trades with positions in other instruments (correlated assets, currencies, etc.). If the primary asset of interest trades in an FBA while the hedges are made in other markets, which are either continuous markets or other FBAs not perfectly synchronized with the primary asset, hedging becomes far riskier, again with the ultimate result of reducing market liquidity and increasing transaction costs. This example vividly illustrates another general criticism of FBAs: unless every single major asset moves to synchronized FBAs (highly unlikely given market, currency, and regulatory fragmentation), there will always be complex games played around latency and order priority.

More recently, Zhang and Ibikunle (2023) found that implementation of the EU’s Markets in Financial Instruments Directive II, which restricts the amount of trading volume which can be executed in so-called ‘dark pools’ and therefore redirects some trading flow to FBA-based alternatives, led to a significant reduction in liquidity associated with this change and an ambiguous effect on overall market quality. They claim that movement to FBA tends to mitigate the degree of “adverse selection,” but they use the same definition of adverse selection costs as in Budish et al., which, as we previously explained, is inherently flawed.

Finally, taking a step back from the gory details of FBAs allows us to see how latency competition emerges naturally from the complexity of real-world markets. Suppose that market prices were purely determined by some kind of divine oracle that emits perfectly accurate price information for all assets every single second at perfectly spaced intervals. In this fantasy world, there would be no need for competition between makers and takers; in fact there would hardly be any need for markets to exist at all. However, prices in the real world obviously do not operate like this. There are tens of thousands of traded assets, all linked together in an extremely complex web of supply and demand and fluctuating in response to the ever-changing state of reality; even if some Platonic ideal of a “true price” exists, there is no way to actually access that information. Instead, the best that we can do is engage in iterative, real-time competition to estimate prices, and the faster this happens, the more price stability we enjoy.

Do FBAs (or other mechanisms of discretizing time or slowing down HFTs) have no place in financial markets? Far from it. According to a 2017 report from the European Central Bank on the prevalence of “dark pools,” in which trades are typically executed in auction formats, up to 15% of trading volume in major names is executed in such pools, suggesting a reasonable amount of natural demand for FBAs. At the same time, if HFT in continuous markets truly imposed significant costs upon all participants, why would the majority of trading volume not shift to dark pools? Clearly, the benefits of continuous markets are still great enough that the lit markets continue to capture the lion’s share of trading activity.

---

The HFT Industry is Small, But Has Positive Externalities

The supposed costs of the HFT industry generate a great deal of consternation. However, in absolute terms, the industry is quite small, and it arguably generates a plethora of positive externalities in a very cost-effective fashion. For example, consider the number of people employed in the industry: Jane Street currently has around 2,600 employees; Citadel has 2,800; Optiver has around two thousand; Virtu Financial and Tower Research have around a thousand. In total, the industry probably employs on the order of 20 to 40 thousand quants, traders, and software engineers worldwide.

This is just not that many people! By comparison, Google alone has nearly two hundred thousand employees; similarly, Meta has over sixty thousand employees. Moreover, even as financial markets grow in size and complexity, the generalizability of trading infrastructure and strategies means that the HFT industry won’t grow nearly as fast, making it a relatively personnel-efficient industry.

For another interesting comparison, let’s quickly look up the total revenue of several major HFT firms:

• Jane Street: $10.6 billion in 2023

• Citadel: $22 billion in 2022 (not all attributable to HFT)

• Virtu Financial: $2.3 billion in 2023

Although revenue data for most firms isn’t available online, Headlands Tech estimates total annual HFT revenue at below $30 billion for most years. Even if we conservatively up that number to, say, $50 billion, it pales in comparison with any slew of commonplace industries. The global tea industry posts revenues exceeding $250 billion; annual toilet paper revenues are >$110 billion; even IKEA has annual revenues exceeding $50 billion. Of course, tea, toilet paper, and IKEA are all reasonably valuable components of daily life, and HFT firms likely face much lower costs by comparison, but stepping back, a $50 billion industry just isn’t all that special in the grand scheme of things.

On the other hand, even though the HFT industry is pretty small by any reasonable metric, one can argue that it generates a disproportionate amount of value for society. For example, consider that the development of infrastructure to support HFT directly spurs the advancement of both hardware (ASICs, FPGAs, specialized processors, low-latency networking, fiber optic networking, data centers) and software (machine learning, satellite image analysis, fast model inference, distributed computing algorithms). If there is any lesson that we’ve learned over the last two decades, it’s that the exponential growth of computing power enables an ever-widening range of heretofore unimaginable applications; in this sense, a source of concentrated demand for high-performance computing yields significant dividends to society at large.

The argument that HFT is a waste of talented labor is also reliant upon a completely inaccurate model of labor productivity and skill. First, HFT is a relatively unique industry, meaning that people who excel in HFT tend to really like the kind of work that they do; even though employees at HFT firms are all undoubtedly talented, it is unclear that they would necessarily have equally high productivity if they were counterfactually allocated to a different segment of the economy. Second, the average tenure at an HFT firm is relatively short, due to high compensation coupled with a stressful work environment. When employees depart for other industries, they take the skills they learned with them and end up cross-pollinating other industries. While it is difficult to precisely quantify the size of this positive externality, the overall argument that HFT serves as a “black hole” that sucks up the output of productive, intelligent employees seems incomplete at best.

In contrast, many discussions of the negative externalities of HFT seem to largely rely on flawed or inaccurate characterizations of their market impacts. For example, consider the oft-repeated argument that high-frequency traders act as “middlemen” who extract value from financial markets, to the detriment of “normal” participants. One obvious gap in this argument is that it has no particular relationship to whether trading is high-frequency or not; if arbitrage is bad, then that applies equally to sub-microsecond arbitrage between ES and SPY as well as to prehistoric merchants who traveled on the Silk Road to take advantage of cross-continent pricing differences.

A more fundamental mistake, however, is that this criticism is really about the allocation of surplus value. Suppose that an uninformed trader, who all informed traders realize is uninformed, executes a large series of market buys and pushes up ES by 1%. If an arbitrageur sells ES (and, optionally, hedges by buying SPY), they capture value generated by that uninformed trade (sort of like pushing a boulder down a hill that someone else pushed up earlier). In the absence of informed arbitrage, an uninformed trader might happen to sell ES as a matter of coincidence and capture some of the surplus which would otherwise have gone to the arbitrageur. However, a number of other outcomes might result: an uninformed trader might buy even more ES, executing trades at a horrible price; alternatively, perhaps SPY traders might inaccurately perceive this as a real movement in the value of the underlying and push up SPY, which might then cause ES traders to push up ES… While simplistic, these examples illustrate that HFT serves to correct and stabilize prices, preventing runaway volatility and giving market participants greater certainty about transacting at fair prices. One could attempt to argue that the value generated (in reducing volatility and spreads, as well as via all other positive externalities) is less than the surplus extracted by HFT firms, and that it would be preferable to return this surplus to the market, but that is a much different argument from claiming that markets would be strictly better off without HFT.

Finally, there are also more speculative arguments that can be made about the positive externalities of HFT. For example, one can also argue that limit order books are the “ideal form” of financial markets insofar as the shape of the supplied liquidity is roughly analogous to the shape of the demand and supply curves for that asset, and that HFT serves as a form of “adversarial computing” which continually updates these demand and supply curves. Alternatively, one can argue that transaction costs are the “gradient signal of the economy,” and consequently that lowering transaction costs in an increasingly financialized world ought to dramatically speed up economic growth. Whether such arguments hold water largely depends on the long-term course of future development, but they are, at the very least, provocative and enticing.

---

Closing Remarks

Ultimately, HFT is not merely a game of speed for its own sake; it is a vital and innovative component of modern financial markets, one that delivers meaningful value by constantly updating prices, tying together disparate markets, and supplying cheap liquidity while driving forward technological progress on many frontiers. It’s easy and naively compelling to imagine high-frequency traders zooming around at lightspeed in the order books, stealing pennies from you every single time you make a trade. In reality, however, HFT firms love nothing more than to give retail traders a good deal—just look at the millions of dollars of price improvement that retail traders have been compensated with in return for their order flow.

At the end of the day, HFT firms aren’t stupid. If it really were the case that competing in the markets requires exponentially accelerating consumption of capital and labor to eke out ever smaller gains in a winner-take-all system, there would be little interest in continuing such destructive competition in the first place. However, this is little more than a simplified and myopic view of the market. Trading firms use a large variety of different strategies: one signal might be only be applicable to a small basket of assets; another signal might be limited to certain segments of time; one might give a continuous readout of potential price movement, whereas yet another might “burst” in response to very specific combinations of market conditions. Assets, exchanges, and regulatory conditions all vary, and no firm comes close to being a sole, dominant player across all markets—hardly a winner-take-all situation.