Zum Inhalt springen

High-Frequency Trading

Zusammenfassung

High-frequency trading is financial speculation conducted by algorithms that hold positions for milliseconds, exploit arbitrage opportunities measured in fractions of a cent, and react to market data faster than human traders can perceive that data exists. What began as academic research into market microstructure in the 1990s grew into an industry that by 2009 accounted for 60–70% of all US equity trading volume, made fortunes for the firms that mastered it, and triggered a public reckoning when the Flash Crash of May 6, 2010, erased nearly $1 trillion in market value in minutes before recovering within an hour. HFT’s story is a collision between financial economics, computer engineering, and public infrastructure — a reminder that when any domain runs on software fast enough, the software becomes the market rather than serving it.

Market Microstructure and the Origins

High-frequency trading grew from academic work in market microstructure — the study of how prices form in financial markets, how buyers and sellers find each other, and how information propagates through trading activity. In the 1980s, economists including Lawrence Glosten, Paul Milgrom, and Maureen O’Hara developed models of how informed and uninformed traders interact through market makers.

The key insight that enabled HFT was about arbitrage and information speed. Financial markets are not perfectly integrated: the same asset trades on multiple exchanges simultaneously, and prices across exchanges can diverge for fractions of a second before arbitrageurs buy on the cheaper exchange and sell on the more expensive one. Anyone who could detect these discrepancies faster and act on them before other traders had a profitable edge.

In the 1990s, this edge was available to any professional with a fast computer and good programming. SOES Bandits (Small Order Execution System Bandits) were retail traders who exploited a NASDAQ system designed to protect small investors — using software to detect and trade on the momentary price discrepancies that human market makers were slow to update. NASDAQ eventually modified SOES to close the loophole, but the bandits demonstrated the principle: computers trading against human reflexes would win.

Decimalization and the Speed Race

The regulatory trigger for HFT was decimalization. In April 2001, US stock markets converted from pricing stocks in fractions (1/8, 1/16 of a dollar) to decimals — penny increments. The minimum price spread between buy and sell prices collapsed from 6.25 cents to 1 cent.

Decimalization was designed to benefit retail investors by reducing the implicit cost of trading. It did — but it also made traditional human market-making economically difficult. When spreads were 6.25 cents, a market maker could quote both the buy and sell price, collect the spread as profit, and absorb inventory risk. At 1-cent spreads, the same strategy yielded far less revenue while the inventory risk remained. Human market makers could not survive at this margin.

Algorithmic market makers could. A computer that could quote thousands of securities simultaneously, adjust its quotes in microseconds as new information arrived, and manage inventory risk through rapid hedging could be profitable at 1-cent spreads. The decimalization transition accelerated the shift from human to algorithmic market-making.

Simultaneously, markets were fragmenting. The Regulation NMS (National Market System) rules, finalized in 2005, required brokers to route orders to the exchange offering the best price nationally, across all exchanges. This created an ecosystem of competing venues — NYSE, NASDAQ, BATS, ARCA, Direct Edge, and dozens of others — all connected by electronic routing. The best-price requirement meant that prices on all venues had to stay close to each other, or arbitrageurs would profit from the discrepancy.

The Technology of Milliseconds

HFT firms built an extraordinary technological infrastructure:

Co-location: HFT firms paid exchanges to place their servers physically inside the exchange’s data center, connected to the matching engine by the shortest possible cable. Light travels 200,000 kilometers per second in fiber optic cable; a co-located server receives market data nanoseconds before a server in a nearby building. Exchanges charged premium fees for co-location because HFT firms valued it enough to pay.

Microwave and laser networks: By 2010, HFT firms were building private microwave transmission networks between exchange data centers because microwave signals travel faster through air than light through fiber optic cable. The Chicago-to-New Jersey route — connecting CME Group’s data center in Aurora, Illinois to NYSE’s data center in Mahwah, New Jersey — was particularly contested. HFT firm Spread Networks spent $300 million in 2010 to build a new fiber route through mountains and bedrock to achieve a round-trip time of 13.1 milliseconds. Within three years, microwave networks had cut the same route to under 9 milliseconds. Eventually, laser links over water were attempted to cut additional fractions of milliseconds.

FPGA trading: Software running on conventional CPUs introduces latency — instructions must be decoded, memory accessed, results written back. Field-Programmable Gate Arrays (FPGAs) implement trading logic directly in reconfigurable hardware, bypassing the software stack. FPGA-based trading systems can receive a market data packet and generate an order response in under a microsecond — too fast for any software-based system to compete.

Kernel bypass: Standard network communication passes through the operating system kernel, adding latency. HFT systems use kernel-bypass networking (DPDK, RDMA, Solarflare OpenOnload) to access the network interface directly from user-space code, removing the kernel overhead.

The Firms and Their Strategies

A handful of firms dominated HFT:

Virtu Financial, founded by Vincent Viola in 2008, was a pure market-maker. Virtu’s 2014 IPO filing revealed that in the five and a half years of data disclosed, Virtu had experienced exactly one losing day. The statistical regularity of its profits illustrated how algorithmic market-making differed from traditional speculation — it was more like collecting a fee on every transaction than betting on market direction.

Citadel Securities (separate from Citadel LLC, the hedge fund) became the largest US equity market maker, handling approximately 25% of US retail equity order flow by 2020. Citadel Securities’ “payment for order flow” arrangement with Robinhood, paying for the right to execute Robinhood’s retail orders, became controversial during the GameStop trading episode of 2021.

Two Sigma, Renaissance Technologies (specifically its Medallion Fund), Jane Street, and Jump Trading operated a mix of HFT and quantitative strategies. Renaissance’s Medallion Fund, founded by mathematician James Simons, returned approximately 66% annually before fees from 1988 to 2018 — the greatest investing track record in history — through quantitative strategies that evolved from academic signal research.

The Flash Crash

On May 6, 2010, beginning at 2:32 PM Eastern time, US equity markets experienced a price collapse of extraordinary speed. The Dow Jones Industrial Average fell nearly 1,000 points — about 9% — in minutes, with individual stocks like Procter & Gamble briefly falling 37% and Accenture trading at $0.01. Then prices recovered, almost entirely, within 36 minutes.

The Flash Crash was caused by an interaction between a large sell order and HFT market-making algorithms. A Kansas City mutual fund, Waddell & Reed Financial, began selling $4.1 billion worth of S&P 500 futures contracts through an automated execution algorithm that was indifferent to price — it would execute the order regardless of market conditions. As the selling pressure increased, HFT market-making algorithms, which had been providing liquidity, detected the unusual conditions and withdrew from the market — their risk models correctly identifying an environment where holding inventory was dangerous.

With liquidity evaporated, prices collapsed as other algorithms triggered further sell orders. The cascade of algorithmic reactions turned a large single trade into a market-wide event visible from any Bloomberg terminal in the world.

The SEC’s investigation, published in October 2010, identified the sequence but concluded that HFT firms’ withdrawal of liquidity was an appropriate response to risk, not a cause. The episode highlighted a structural vulnerability: a market that depended on algorithmic market-makers for liquidity could see that liquidity disappear instantly under stress, precisely when it was most needed.

Michael Lewis and the Public Reckoning

Public awareness of HFT remained limited until Michael Lewis published Flash Boys in March 2014. Lewis’s narrative centered on Brad Katsuyama, a Royal Bank of Canada trader who had noticed that his large buy orders moved markets against him before he could execute them. Katsuyama’s investigation revealed that HFT firms were using their speed advantage to detect his intent and trade ahead of his orders — a practice called “front-running.”

Lewis’s framing — that HFT firms were front-running ordinary investors, that “the market is rigged” — was explosive. The book became a bestseller; Congressional hearings followed; the FBI opened an investigation; the New York attorney general launched a probe into exchanges.

The financial economics community largely disagreed with Lewis’s framing. Academic research consistently found that HFT had reduced bid-ask spreads and increased the immediacy of order execution for ordinary investors. HFT’s presence meant that retail investors paid lower transaction costs than in the era of human market-makers. The “front-running” Lewis described was not illegal; it was a consequence of how equity markets were structured.

The debate was never fully resolved. Lewis was not wrong that HFT firms profited by being faster; the economists were not wrong that the overall cost of trading had fallen. Both things could be true simultaneously: a system can lower average trading costs while allowing sophisticated players to extract rents from informational speed advantages.

Regulation and Normalization

Regulators in the US and Europe moved to address the most destabilizing aspects of HFT without prohibiting it entirely. Market-wide circuit breakers, introduced after the Flash Crash, pause trading automatically when prices move beyond defined thresholds. Limit up-limit down rules prevent individual stocks from trading outside a price band. Consolidated Audit Trail (CAT) requirements improved regulators’ ability to reconstruct events after they occurred.

The HFT industry itself consolidated. Early HFT was profitable because the firms that mastered it were rare; as the strategies became understood and the technology commoditized, returns compressed. By the mid-2010s, HFT’s share of US equity volume had fallen from 60% to roughly 50%, as less profitable players exited and the remaining firms competed on ever-smaller margins.

HFT remains a defining feature of modern financial markets — a demonstration that when infrastructure is fully digitized and competitive, speed becomes a resource as valuable as capital, and the race to exploit it reshapes the infrastructure itself.

📚 Sources