High-Frequency Trading and Neural Networks: The Nanosecond Arms Race

In 2010 a company called Spread Networks spent an estimated $300 million on a nearly straight fiber optic cable between New York and Chicago, cutting the round-trip signal time from 16 milliseconds to 13. Traders called three milliseconds an eternity, and within a few years microwave links had cut the trip to 8.5 milliseconds. That is the world of high-frequency trading (HFT), where firms use computers to buy and sell in fractions of a second. Neural networks, which are software models that learn patterns from data, are now one of the tools used to predict price moves inside that window.

Educational only. Not financial advice. Investing involves risk of loss; past performance does not guarantee future results. This article explains how a segment of the market works and does not recommend any trading approach or security.

The central tension is simple. Speed is a physical resource, and physics sets a floor under it. Light travels about 30 centimeters (1 foot) in a nanosecond in a vacuum and roughly one-third slower in glass fiber, so a firm's distance from an exchange is part of its business model. Prediction is the other resource, and that is where machine learning comes in. This article walks through how the two combine, what published research can and cannot say about neural networks in this setting, and what the evidence shows about whether HFT helps or harms ordinary investors.

Key takeaways

• Estimates cited at a June 2010 SEC roundtable put HFT at "50 percent or more" of U.S. equity trading volume, though the number depends on how HFT is defined and has been debated ever since.
• A 2015 study found that the duration of one price-gap opportunity shrank from a median of 97 milliseconds in 2005 to 7 milliseconds in 2011, while the profit per trade stayed roughly constant.
• Deep learning models such as DeepLOB predict short-term price moves from the order book, but a 2024 preprint found that high forecasting power "does not necessarily correspond to actionable trading signals."
• Research by Hasbrouck and Saar found that more low-latency activity was associated with narrower spreads, deeper order books, and lower short-term volatility. Other research argues the speed race itself is wasteful.
• In the 2010 Flash Crash, automated traders, including high-frequency firms, pulled back or passed contracts among themselves while prices fell fast. The SEC and CFTC staff traced the trigger to a single large sell order, not to HFT alone.

Why a millionth of a second can be worth money

Most trading profits come from knowing something slightly before everyone else. For a long-term investor that something may be a company's earnings, and the advantage lasts days. For an HFT firm it may be that the price of one product just changed and a related product has not caught up, and the advantage lasts milliseconds.

The clearest published measurement comes from economists Eric Budish, Peter Cramton, and John Shim, who studied two products that track the S&P 500 index: the E-mini futures contract (ticker ES, traded in Chicago) and the SPDR exchange-traded fund (ticker SPY, traded in New Jersey). Over a day or an hour the two move almost in lockstep. At very short intervals they do not: the median correlation of their returns over 2011 was 0.0923 at 10 milliseconds and 0.0073 at 1 millisecond, which is close to zero. In effect, the two prices are the same story told by two narrators who hear the news at slightly different moments.

That gap creates a temporary profit opportunity each time one price jumps before the other. Budish and colleagues calculated about a thousand such opportunities per day in ES and SPY, worth on the order of $75 million per year. They found that the duration of these opportunities fell from a median of 97 milliseconds in 2005 to 7 milliseconds in 2011, yet the profitability per trade stayed remarkably constant at a median of about 0.08 index points. Their interpretation is that the speed race does not eliminate the opportunity. It just raises the bar for capturing it (Budish, Cramton, and Shim, working paper, 2013; published in the Quarterly Journal of Economics, 2015).

Think of it as a game of musical chairs where the music keeps getting shorter. The number of chairs stays the same, but the players have to invest ever more in reflexes to find one.

The physics of the race

The table puts the scales in perspective. These numbers describe different things (human reaction, network paths, and deliberate delays), but together they show where the competition lives.

Event	Time	Source
Light travel, 1 nanosecond	About 30 cm in a vacuum	Physical constant
Intentional delay added by the IEX exchange to each message	350 microseconds	SEC approval coverage, June 2016
New York to Chicago round trip, older fiber routes	16 milliseconds	Budish et al.
New York to Chicago round trip, Spread Networks fiber (2010)	13 milliseconds	Budish et al.
New York to Chicago round trip, microwave (by about 2013)	8.5 milliseconds	Budish et al.
Light, one way between Chicago and New York	About 4 milliseconds	Budish et al.
Blink of a human eye	400 milliseconds	Budish et al.

A microsecond is a thousandth of a millisecond, and a nanosecond is a thousandth of a microsecond. Budish and colleagues note that the E-mini and SPY correlation must account for the roughly 4 milliseconds that light needs to travel between Chicago and New York. That is not a figure of speech: special relativity places a floor on how fast information can flow between two markets.

For the engineering behind the glass lines carrying these signals, see How Fiber Optic Cables Transmit Data.

Where the money comes from: four strategies

HFT is not a single strategy. It is a set of strategies that share a need for speed. Large firms in this space include Citadel Securities, Virtu Financial, Jane Street, Jump Trading, and Hudson River Trading, and most of the largest are private, so little of their financial detail is public. The four broad families are listed below.

Market making. A market maker posts both a price at which it will buy (the bid) and a price at which it will sell (the ask). The difference is the spread, and the maker earns it when it trades with buyers and sellers in turn. The risk is inventory: if everyone sells to you and the price keeps falling, you lose. Short-term prediction helps by telling the maker when to widen or tighten its quotes.

Statistical arbitrage. This strategy bets that two related prices, such as the same stock on two venues, or an index and its components, will return to their usual relationship. The ES and SPY example above is a case. Speed matters because other firms notice the same gap.

Latency arbitrage. A firm that sees a price change on one venue first can trade against slower quotes on another before they update. Michael Lewis's 2014 book Flash Boys popularized the practice. It is legal in the U.S. and also a source of ongoing controversy because the profit comes from the difference in speed and not from new information about value.

Short-horizon prediction. The fourth family uses models to forecast price direction over seconds or fewer from the state of the order book and recent trades. This is where neural networks have the most direct application, and it is the one this article spends the most time on.

Each strategy earns a tiny amount per trade on a very large number of trades. A firm that makes a fraction of a cent per share needs to be right slightly more often than wrong, over millions of repetitions. That structure also explains why transaction costs, latency, and even the placement of a server rack matter so much.

For a related look at how machine learning is used across trading, see Algorithmic Trading and Neural Networks, and for the market-liquidity concept that underpins market making, see Why Liquidity Matters in Markets.

Neural networks and the limit order book

The limit order book is the live list of all outstanding offers to buy and sell a security at each price. Picture a stack of buy orders sitting below the current price and a stack of sell orders sitting above it. When a buyer is willing to pay the lowest asking price, a trade occurs and the stacks change. A busy stock's book updates thousands of times per second.

The book carries information. If the stack of buy orders is much larger than the stack of sell orders, price is slightly more likely to tick up next. Older models measured this with hand-built features such as order imbalance (the ratio of buy to sell volume at the top levels). Neural networks learn such features directly from raw data. If hand-built features are like a chef following a recipe, a neural network is like a chef who tastes thousands of dishes and works out the recipe on their own. For readers who want the mechanics of the models themselves, How Do Neural Networks Work? covers the basics.

DeepLOB and what it showed

A widely cited paper is "DeepLOB: Deep Convolutional Neural Networks for Limit Order Books" by Zihao Zhang, Stefan Zohren, and Stephen Roberts of the Oxford-Man Institute of Quantitative Finance at the University of Oxford, published in IEEE Transactions on Signal Processing in June 2019. The model stacks two kinds of layers. Convolutional layers scan the book for local patterns, such as the shape of the queues at neighboring price levels. A recurrent layer called an LSTM (long short-term memory) then tracks how those patterns develop over time. The authors report that the model beat earlier methods on a benchmark limit order book dataset and that it kept stable accuracy out of sample on a year of London Stock Exchange quotes, including on instruments it was not trained on. They read that as evidence that the network learns features common across stocks (Zhang, Zohren, and Roberts, arXiv:1808.03668).

What the public benchmark covers, and what it does not

Academic work on order book prediction needs shared data, and exchange data is expensive. The best-known public answer is FI-2010, a benchmark released by Adamantios Ntakaris and colleagues. It holds roughly 4 million time-series samples from five stocks on the Nasdaq Nordic market over 10 consecutive trading days, together with a day-based cross-validation protocol and baseline results (Ntakaris et al., arXiv:1705.03233). The data let researchers compare models on equal terms, which is the job of any benchmark.

The size of the data also sets its limits. Five stocks and two trading weeks cannot represent thousands of listed securities across calm and violent markets, and a model that wins on this slice may or may not survive elsewhere. That is why the London Stock Exchange test in the DeepLOB paper, a full year of quotes with instruments held out of training, carries more weight than a leaderboard score. It is also why the next question matters more than the next accuracy point.

Why prediction accuracy is not the same as profit

A more recent paper pushes back on how such results should be judged. Antonio Briola, Silvia Bartolucci, and Tomaso Aste studied a varied set of NASDAQ stocks and released an open-source toolkit called LOBFrame. They found that a stock's microstructure affects how well deep learning works on it, that high forecasting power does not necessarily correspond to actionable trading signals, and that traditional machine learning metrics fail to adequately assess forecast quality in this setting. They propose judging forecasts by how well they predict complete transactions (Briola, Bartolucci, and Aste, arXiv:2403.09267, preprint first submitted March 2024).

The gap between accuracy and profit has several causes that practitioners recognize. A model can be right about the direction of a one-tick move that is too small to cover the exchange fee and the spread. A model's forecast can be right but arrive after the price has already moved. And a model that is accurate at predicting something frequent but unprofitable, such as "the price will not change," will score well and earn nothing. These are the same traps covered in Why 99% Accuracy Can Be Useless.

Why markets break models

Financial data differ from the photographs that image models learn on. The statistical pattern that holds today can fade tomorrow because other firms trade on the same pattern, a process that destroys the edge. A model trained on a calm month may behave strangely in a volatile one. This is the problem described in What Model Drift Means in Production and in Why Overfitting Hurts Machine Learning, only more pressing because the environment fights back.

Hardware: the model must fit the time budget

A neural network that predicts well in 50 milliseconds is worthless if the opportunity lasts 7. That constraint changes how the models are built. In most machine learning, bigger is better and the hardware is a cluster of GPUs (graphics processing units, chips that excel at parallel arithmetic; see How GPUs Work). In HFT, the model must answer within a budget set by the market.

Firms therefore trim and shrink models. Common techniques include using fewer layers, reducing numerical precision (so that each number occupies fewer bits), and compiling the model onto an FPGA (field-programmable gate array), a chip whose circuits can be rewired to implement a specific calculation without an operating system in the way. Specialized chips, called ASICs (application-specific integrated circuits), take the idea further by hardwiring one design. The exact deployments are trade secrets, and published latency figures for neural networks on real order books are scarce, so any claim of a specific "nanosecond neural network" should be treated with caution.

The practical result is a trade-off between accuracy and speed. A giant language model that can write an essay is useless here, and a small model that answers in a microsecond may be worth more than a larger one that answers in a millisecond. See Why AI Needs So Much Computing Power for the opposite end of the trade-off.

Design choice	What it buys	What it costs
Smaller network	Faster inference	Less capacity to learn complex patterns
Lower numerical precision	Faster arithmetic, less memory	Some accuracy loss
FPGA implementation	Predictable low latency	Long development cycle, hard to update
Custom ASIC	Fastest possible for one design	Very high cost, no flexibility
Co-location near the exchange	Shorter signal path	Fees, rack space, competition for position

Does HFT help or hurt? What the research says

The evidence is mixed, and the honest summary is that different studies answer different questions.

Evidence for. Joel Hasbrouck and Gideon Saar built a measure of low-latency activity from public NASDAQ data. They found that increased low-latency activity improves traditional market quality measures, namely decreasing spreads, increasing displayed depth in the limit order book, and lowering short-term volatility. They concluded that, under the current structure of U.S. equity markets, more low-latency activity need not harm long-term investors (Hasbrouck and Saar, Journal of Financial Markets, 2013).

Attribution is tricky. Quoted spreads in U.S. stocks fell sharply when exchanges switched from fractions to one-cent prices (decimalization), which was completed on April 9, 2001, years before HFT reached its present scale. SEC staff testimony reported that quotation spreads in NYSE-listed securities narrowed about 37 percent on average and Nasdaq spreads about 50 percent, while quote sizes shrank by 60 percent on the NYSE (SEC testimony, May 2001). Much of the narrowing in spreads therefore came from tick size and competition, and it is hard to separate from the effect of faster traders.

Evidence against. Budish and colleagues argue that the continuous limit order book is a flawed market design. The speed race, in their model, is a prisoner's dilemma: each firm must spend on speed to avoid losing the prize to rivals, yet the prize is not new information, so the spending is socially wasteful, and the cost is ultimately borne by investors through wider spreads and thinner markets. Their proposed remedy, frequent batch auctions, would collect orders over a short interval (one second in the working paper) and match them all at one price, so a microsecond advantage no longer wins (Budish, Cramton, and Shim). The logic of the arms race is explained in Game Theory and the Nash Equilibrium.

A market-design response. The exchange IEX took a different route. It adds a 350-microsecond delay to incoming and outgoing messages, a "speed bump" meant to blunt latency arbitrage. The SEC approved IEX as a national securities exchange on June 17, 2016, with one commissioner dissenting on parts of the decision, concluding that the delay was small enough not to prevent investors from accessing prices fairly and efficiently, and it said it would study the effects. Critics such as Nasdaq argued the delay conflicted with the rule that prices be available to all at the same time.

Both pieces of research can be true at once. A market can offer tighter quotes in normal times and still be an expensive speed contest that benefits a small group of firms, and neither fact settles whether regulation should change.

Counting the race: what exchange message data show

The 2015 Budish paper measured the race in two index products. A later study went after the whole market. Matteo Aquilina, Eric Budish, and Peter O'Neill used message data from the London Stock Exchange, which records every attempt to trade or cancel, including the attempts that lose. Ordinary order book data shows only the winners. The sample covered all FTSE 350 stocks over nine weeks in fall 2015, and the paper was published in the Quarterly Journal of Economics in 2022 (Harvard Law School Forum summary; FCA Occasional Paper 50).

In this research a race is a burst of competing messages aimed at the same price, such as several firms trying to take one stale quote before anyone else does. The headline measurements are in the table.

Measure	Finding (FTSE 100 stocks, fall 2015)
Races per symbol per day	About 537, roughly one per minute
Typical gap between winner and first loser	5 to 10 microseconds
Share of trading volume in races	About 22 percent of volume, 21 percent of trades
Concentration	The top six firms win about 82 percent of races and lose about 87 percent
Implied cost to investors	About 0.42 basis points of total volume (1 basis point is 0.01 percent)
Effect of removing races, for example with batch auctions	Roughly 17 percent lower cost of liquidity
Extrapolated annual size	About GBP 60 million in the UK and about $5 billion across global equities

Source: Aquilina, Budish, and O'Neill, as summarized by the Harvard Law School Forum on Corporate Governance, November 2021. The global figure is an extrapolation by the authors.

Two details stand out. First, the prize per race is small, which is consistent with the earlier ES and SPY finding that profit per opportunity stays roughly constant while the window shrinks. Second, a handful of firms win and lose nearly all the races, which is the signature of a market where the entry ticket is infrastructure rather than insight. The same six firms take about 80 percent of liquidity in races while providing about 42 percent of the liquidity taken, so in these races the leading firms are net takers of liquidity rather than providers of it.

An illustrative calculation shows why a fraction of a cent is enough. On a $50 stock with a one-cent tick, half a tick is $0.005 per share, or 1 basis point of the price. If a firm wins that half tick across 10,000 shares, it earns $50 before costs. That sum is trivial for one trade and large across hundreds of races per symbol per day, which is how a business can be built on a gap no human could see. The example is arithmetic, not a claim about any firm's profits.

The rules that shape the race

Speed contests do not happen in a legal vacuum. In the United States, Regulation NMS (National Market System), adopted in 2005, ties the venues together. Its centerpiece, Rule 611 or the order protection rule, requires trading centers to keep policies that prevent a trade-through, which means executing an order at a price worse than a better protected quote shown on another exchange. The rule covers only the top of the book as shown in the consolidated data feed, and the 2005 adoption was a 3 to 2 vote of the SEC commissioners (WilmerHale, June 2026).

The rule helps explain why latency matters so much. If a better price exists on another venue, a trader must route there, so every venue's best quote becomes a shared target that fast firms watch across all of them. Critics argue the same structure encourages fragmentation and a proliferation of order types. On June 11, 2026, the SEC voted to propose rescinding Rule 611 and the related Rule 610(e) ban on locked and crossed markets. As of that report this was a proposal, not a final rule, with comments due 60 days after publication in the Federal Register, so the rule remains in force until the Commission acts.

When automated liquidity disappears: the 2010 Flash Crash

On May 6, 2010, major U.S. stock indices that were already down more than 4 percent suddenly fell a further 5 to 6 percent within minutes and then rebounded almost as fast. Over 20,000 trades across more than 300 securities were executed at prices more than 60 percent away from their values moments earlier, some at a penny and some as high as $100,000. By the close, major indices were down about 3 percent from the previous day.

The joint SEC and CFTC staff report attributes the trigger to a single large trader. At 2:32 p.m., a mutual fund complex began selling 75,000 E-mini contracts (about $4.1 billion) through an automated algorithm that targeted 9 percent of trading volume over the previous minute, without regard to price or time. The program was executed in 20 minutes, while a similar sale earlier had taken more than five hours.

The report describes the behavior of high-frequency firms in specific terms. High-frequency traders first absorbed the selling and built a net long position of about 3,300 contracts. Between 2:41 and 2:44 p.m., they sold roughly 2,000 contracts to reduce it while trading nearly 140,000 contracts, over 33 percent of volume. Then, between 2:45:13 and 2:45:27, they traded more than 27,000 contracts, about 49 percent of volume, while buying only about 200 additional contracts net, a "hot potato" effect as positions passed back and forth. Buy-side depth in the E-mini fell to less than 1 percent of the morning's level. At 2:45:28 p.m., the CME's stop logic paused E-mini trading for five seconds, after which prices stabilized (SEC and CFTC staff, "Findings Regarding the Market Events of May 6, 2010," September 30, 2010).

The report is also clear that a second liquidity crisis occurred in individual stocks, where many liquidity providers paused their automated systems after the first drop, and some widened quotes or withdrew entirely. Automated market makers have no obligation to keep trading when conditions look dangerous, and their pause is what turned a sharp decline into a vacuum.

Speed is not the only way to lose money fast

A second cautionary case does not involve any neural network. On August 1, 2012, a software installation problem at Knight Capital Group caused it to send numerous erroneous orders into NYSE-listed securities at the market open. Knight reported a pre-tax loss of about $440 million, and on August 6 it agreed to sell $400 million of convertible preferred stock to raise capital (Knight Capital Form 8-K, August 2012). The lesson for AI-driven trading is that automated systems fail at machine speed, and risk controls have to operate at the same speed.

Could neural networks make this worse?

A hypothesis is that many firms training similar models on similar data might react to stress in similar ways, so that liquidity would withdraw in unison. That is plausible reasoning, but it is a hypothesis. I found no published study that measures it for neural-network-driven HFT, and the 2010 report describes the behavior of rule-based algorithms that long predate today's deep learning. A fair reading is that the risk deserves monitoring and that evidence for it is limited.

Who earns the profits

Because most large HFT firms are private, audited numbers are rare. Virtu Financial, which is public, reported consolidated trading income of about $1.82 billion for 2024, up about 40 percent from roughly $1.30 billion in 2023, on total revenues of about $2.88 billion (Virtu Financial Form 10-K for fiscal 2024). Revenue figures reported for private firms such as Jane Street and Citadel Securities in trade press come from documents seen by reporters and are not audited public filings, so this article does not cite them.

Two points matter here. First, revenue is not profit, because firms spend heavily on technology, data, and exchange fees. Second, trading profits are not a measure of social value. A firm can profit by providing liquidity that investors value, by capturing a speed advantage that investors pay for, or by both.

What people get wrong about HFT and AI

Common belief	What the evidence supports
"HFT firms predict the market using AI"	They predict very short-term moves, mostly from the order book. Long-term prediction is a different, much harder problem.
"Neural networks made HFT possible"	HFT existed with hand-built models. Neural networks are a newer tool whose practical edge is debated.
"Spreads narrowed because of HFT"	Much of the narrowing followed decimalization in 2001. Hasbrouck and Saar find low-latency activity is associated with narrower spreads, but attribution is difficult.
"HFT caused the 2010 Flash Crash"	The SEC and CFTC staff traced the trigger to one large algorithmic sell order, with automated liquidity providers amplifying it.
"The fastest firm always wins"	Opportunities shrink as speed rises, but profit per opportunity stayed roughly constant in the Budish data. Speed is a cost of staying in the game.
"Retail investors lose to HFT on every trade"	The evidence is mixed and depends on the measure. Budish and colleagues argue that costs of the speed race fall on investors through wider spreads, while Hasbrouck and Saar find narrower spreads with more low-latency activity.

Frequently asked questions

Q: How do high-frequency firms use neural networks differently from other AI applications? A: Latency is the binding constraint. A model must produce a prediction before the opportunity vanishes, so firms use small networks, reduced numerical precision, and dedicated hardware such as FPGAs. Most are far smaller than the large language models in the news, because inference speed matters more than raw capability.

Q: Do neural networks reliably predict short-term prices? A: Not reliably. DeepLOB reported strong results on benchmark data and on a year of London Stock Exchange quotes. A 2024 preprint studying NASDAQ stocks found that high forecasting power does not necessarily correspond to actionable trading signals, and that results depend on each stock's microstructure.

Q: What is latency arbitrage, and is it legal? A: It is trading against stale quotes. A fast firm sees a price change on one venue and trades on another before slower participants update their quotes. It is legal in the U.S., and critics argue it transfers money from slower traders to faster ones without improving price discovery. IEX's 350-microsecond delay was designed to blunt it.

Q: What caused the 2010 Flash Crash? A: The SEC and CFTC staff report traced the trigger to a large automated sale of 75,000 E-mini contracts (about $4.1 billion) executed in 20 minutes at a fixed share of volume. High-frequency firms and other liquidity providers first absorbed the selling, then reduced positions and passed contracts among themselves, and liquidity thinned in both the E-mini and individual stocks.

Q: Can individual investors compete on speed? A: No. A long-term investor is playing a different game, since price changes over days or years depend on earnings and economic conditions rather than microseconds. Research by Hasbrouck and Saar suggests that increased low-latency activity need not harm long-term investors under current U.S. equity market structure, though the evidence is not unanimous.

Q: How big is HFT in U.S. stock trading? A: At a June 2010 SEC roundtable, a moderator said estimates varied widely but some believed high-frequency traders were 50 percent or more of volume. The figure depends on the definition used and on whether volume is counted once or on both sides, so it should be read as a rough order of magnitude.

Q: How much of trading is a pure speed race? A: One study of London Stock Exchange message data from fall 2015 found that races for stale quotes accounted for about 22 percent of FTSE 100 trading volume, with the typical winner beating the first loser by 5 to 10 microseconds. The authors estimated the cost to investors at about 0.42 basis points of volume. These figures describe one exchange and one period, so they are not a universal constant.

The takeaway

High-frequency trading is a business built on tiny gaps that close in milliseconds, and the evidence suggests the gaps do not go away as firms get faster, they just get captured sooner. Neural networks fit this business well in principle, because the limit order book is a rich stream of data, yet the best recent research warns that forecasting skill does not automatically become profit. Whether the whole arrangement serves investors is a market-design question, with studies pointing both ways. For a reader trying to separate hype from fact, the most useful habit is to ask three things of any claim about AI trading: what time horizon, measured how, and after what costs.

Sources

• Budish, Cramton, and Shim, "The High-Frequency Trading Arms Race: Frequent Batch Auctions as a Market Design Response" (working paper, 2013; Quarterly Journal of Economics, 2015).
• SEC and CFTC staff, "Findings Regarding the Market Events of May 6, 2010", September 30, 2010.
• Zhang, Zohren, and Roberts, "DeepLOB: Deep Convolutional Neural Networks for Limit Order Books", IEEE Transactions on Signal Processing, 2019.
• Briola, Bartolucci, and Aste, "Deep Limit Order Book Forecasting", arXiv preprint, 2024.
• Hasbrouck and Saar, "Low-Latency Trading", Journal of Financial Markets, 2013.
• SEC, "Testimony on decimalization", May 2001, and roundtable transcript on market structure, June 2, 2010.
• Aquilina, Budish, and O'Neill, "Quantifying the High-Frequency Trading Arms Race," Quarterly Journal of Economics, 2022, via the Harvard Law School Forum summary and FCA Occasional Paper 50.
• Ntakaris et al., "Benchmark Dataset for Mid-Price Forecasting of Limit Order Book Data with Machine Learning Methods", Journal of Forecasting.
• WilmerHale, "The SEC Takes Aim at the Trade-Through Rule", June 2026.
• Virtu Financial, Form 10-K for fiscal 2024.

---

Read next:

• Algorithmic Trading & Neural Networks: How AI Makes Billions Per Second
• Crypto Trading Bots and Machine Learning: Separating Hype from Reality
• Market Sentiment Analysis: How NLP Models Read the Market's Mind