TL;DR

High-frequency trading, or HFT, is not one strategy. It is better understood as a bundle of low-latency trading activities that share the same infrastructure but have different social value: electronic market-making can improve liquidity, while latency arbitrage, queue races, spoofing, and manipulative order-book games range from contested to illegal.

The academic evidence is mixed but not random. Hendershott, Jones, and Menkveld find that algorithmic trading narrowed spreads and reduced adverse selection for large-cap stocks; Brogaard, Hendershott, and Riordan find that HFTs contribute to price discovery by trading with permanent price changes and against transitory errors; Hasbrouck and Saar find low-latency activity associated with lower spreads, deeper displayed books, and lower short-term volatility. (faculty.haas.berkeley.edu)

The strongest critique is not that “speed is bad.” It is that the continuous limit order book creates a winner-takes-all race where tiny speed advantages capture stale quotes. Budish, Cramton, and Shim argue that frequent batch auctions could transform competition from speed to price. (OUP Academic)

1. What HFT actually is

HFT is the use of automated strategies, low-latency infrastructure, and direct market connectivity to submit, cancel, and modify orders at very high speed. The key unit is not “the trade” but the order-book event: quote update, order insertion, cancellation, partial fill, full fill, reject, or routing decision.

A useful decomposition is:

  1. Market-making: continuously quote bid and ask prices, earn spread and rebates, manage inventory and adverse selection.
  2. Statistical arbitrage: trade short-horizon mean reversion, lead-lag, ETF-basket, futures-cash, or cross-venue relationships.
  3. Latency arbitrage: race to react to public information before stale quotes are cancelled.
  4. Predatory or manipulative tactics: spoofing, layering, quote stuffing, and momentum ignition.

The first category is often liquidity-providing. The third is legally permitted in many contexts but socially contested. The fourth can cross into illegal manipulation, especially spoofing, which is covered by the U.S. Commodity Exchange Act anti-spoofing provision. (法律資訊學院)

2. The limit order book

A limit order book, or LOB, is a queue of buy and sell limit orders sorted by price and priority. The best bid is the highest resting buy price; the best ask is the lowest resting sell price. The spread is:

\[\text{Spread} = P_{\text{ask}} - P_{\text{bid}}\]

The midprice is:

\[M = \frac{P_{\text{bid}} + P_{\text{ask}}}{2}\]

Most equity venues use price-time priority: better price first, then earlier time at the same price. That makes queue position valuable. A passive market maker is not only choosing a price; it is buying or losing a place in the queue.

A market order guarantees immediacy but not price. A limit order guarantees price but not execution. An immediate-or-cancel order fills whatever is available now and cancels the rest. Hidden and iceberg orders reduce information leakage but usually sacrifice some priority.

3. Why latency matters

In a continuous market, time priority converts microseconds into economics. If a futures price moves first, then ETF quotes, stock quotes, options quotes, and correlated venues may become stale for a tiny interval. The fastest participant can trade against those stale quotes before slower market makers cancel or reprice.

This is the mechanical race emphasized by Budish, Cramton, and Shim. They argue that the continuous limit order book creates frequent latency-arbitrage opportunities and that faster technology reduces the duration of each opportunity without necessarily eliminating the rent. (OUP Academic)

The infrastructure stack usually includes:

  • colocation near exchange matching engines;
  • direct data feeds instead of slower consolidated feeds;
  • kernel bypass and optimized networking;
  • FPGA or hardware-accelerated market-data parsing;
  • microwave or millimeter-wave links where geography matters.

The economic point is simple:

\[\text{Private value of speed} = - \text{cost of speed infrastructure}\]

The social question is harder: does the speed investment improve price discovery enough to justify the arms race, or does it mostly transfer value from slower liquidity providers and end investors?

4. Order lifecycle

A simplified order lifecycle is:

  1. Strategy decides to quote or trade.
  2. Internal risk checks approve or reject the order.
  3. Order is sent through a gateway.
  4. Exchange sequencer timestamps and orders messages.
  5. Matching engine applies book priority.
  6. Execution, partial execution, cancellation, or rejection is returned.
  7. Clearing and settlement follow after execution.

For U.S. equities, the standard settlement cycle moved to T+1 for applicable securities transactions on May 28, 2024. (投資者.gov)

5. The economics of market-making

A market maker posts a bid and an ask. Its rough expected P&L is:

\[\text{P\&L} \approx (\text{spread captured} \times \text{volume}) + \text{rebates} - \text{adverse selection} - \text{inventory cost} - \text{fees}\]

The market maker wants fills, but not toxic fills. It wants volume, but not inventory blowups. It wants queue priority, but not quotes that get picked off when the market moves.

The core risks are:

Inventory risk: the maker accumulates a position and the market moves against it.

Adverse selection: the maker trades with someone who knows more or reacts faster.

Queue risk: the maker is too far back in the queue to get profitable fills but still exposed when the price moves.

Latency risk: the maker cannot cancel stale quotes fast enough.

6. Avellaneda–Stoikov market-making model

The Avellaneda–Stoikov model is a canonical baseline for optimal high-frequency market-making. It assumes the midprice follows a diffusion, order arrivals are stochastic, and the dealer maximizes expected utility under inventory risk. The model studies a dealer submitting bid and ask quotes in a limit order book with inventory and transaction risk. (Cornell ORIE)

Let:

  • \(s\) be the midprice,
  • \(q\) be inventory,
  • \(\gamma\) be risk aversion,
  • \(\sigma\) be volatility,
  • \(T-t\) be remaining horizon.

The reservation price is:

\[r(s,q,t) = s - q\gamma\sigma^2(T-t)\]

If the maker is long, \(q>0\), the reservation price moves lower, encouraging selling. If the maker is short, \(q<0\), it moves higher, encouraging buying.

A common closed-form expression for the total optimal spread is:

\[\delta_a + \delta_b = \gamma\sigma^2(T-t) + \frac{2}{\gamma}\ln\left(1+\frac{\gamma}{k}\right)\]

where order arrival intensity is often modeled as:

\[\lambda(\delta) = A e^{-k\delta}\]

The first term prices inventory and volatility risk. The second term prices the tradeoff between fill probability and per-trade edge.

In practice, this model is not enough by itself. Real systems add queue modeling, adverse-selection filters, volatility regime detection, venue selection, fee logic, hedge logic, and kill switches.

7. Why spreads exist: adverse selection

The Glosten–Milgrom intuition is that a market maker faces informed and uninformed traders. A buy order is more likely when the asset is worth more than the current quote, so the ask must be higher than the unconditional expected value. A sell order is more likely when the asset is worth less, so the bid must be lower.

In compact form:

\[A = \mathbb{E}[V \mid \text{buy}]\] \[B = \mathbb{E}[V \mid \text{sell}]\]

The spread compensates the market maker for losing to informed flow:

\[\text{Spread} = A - B\]

Kyle’s model expresses price impact as:

\[P = P_0 + \lambda y\]

where \(y\) is aggregate order flow and \(\lambda\) is Kyle’s lambda, a measure of illiquidity or price impact.

Roll’s spread estimator uses bid-ask bounce:

\[\text{Spread} = 2\sqrt{-\operatorname{Cov}(\Delta p_t,\Delta p_{t-1})}\]

These old models remain useful because HFT compresses the same economics into smaller time units.

8. Core HFT signals

Order-book imbalance

A simple imbalance signal is:

\[I = \frac{V_{\text{bid}} - V_{\text{ask}}} {V_{\text{bid}} + V_{\text{ask}}}\]

If bid-side depth is much larger than ask-side depth, the next short-term price move is more likely to be upward, all else equal. In production, firms rarely use raw top-of-book imbalance alone. They decay it, normalize it, condition it on spread and volatility, and measure how predictive it is after fees and queue position.

Microprice

The microprice shifts the midprice toward the side with weaker liquidity:

\[\text{Microprice} = \frac{ P_{\text{ask}}V_{\text{bid}} + P_{\text{bid}}V_{\text{ask}} }{ V_{\text{bid}} + V_{\text{ask}} }\]

If there is heavy bid depth and thin ask depth, the microprice moves toward the ask, indicating upward pressure.

VPIN and toxicity

VPIN, or volume-synchronized probability of informed trading, estimates order-flow toxicity using volume imbalance in equal-volume buckets. Easley, López de Prado, and O’Hara introduce VPIN as a toxicity metric updated in volume time. (SSRN)

A simplified expression is:

\[\operatorname{VPIN} = \frac{ \sum_{\tau=1}^{n} \left|V_{\tau}^{S} - V_{\tau}^{B}\right| }{ nV }\]

where \(V_{\tau}^{S}\) and \(V_{\tau}^{B}\) are sell and buy volume estimates in bucket \(\tau\), and \(V\) is bucket size.

High toxicity means liquidity provision is dangerous: quotes are more likely to be hit just before the price moves against them.

9. Latency arbitrage

Latency arbitrage happens when one venue or instrument reflects new information before another. Example:

  1. E-mini S&P 500 futures move.
  2. SPY or related ETF quotes lag for milliseconds.
  3. The fastest trader buys stale asks or sells stale bids.
  4. Slower market makers cancel too late.
  5. Market makers widen future quotes to compensate.

Budish, Cramton, and Shim’s proposed fix is frequent batch auctions: instead of processing orders continuously one by one, process them in discrete batches at a uniform clearing price. Their argument is that discrete time reduces the value of being infinitesimally faster and shifts competition toward price. (OUP Academic)

In a continuous market:

\[\text{Winner} = \arg\min_i(\text{latency}_i)\]

In a batch auction, orders inside the same batch compete primarily on price, not nanosecond arrival time.

10. Statistical arbitrage and cross-asset HFT

Statistical arbitrage is not unique to HFT, but HFT applies it at shorter horizons.

Common structures include:

  • futures versus ETF;
  • ETF versus basket;
  • ADR versus local listing;
  • cash index versus futures;
  • spot versus perpetual futures;
  • options versus underlying;
  • cross-exchange crypto pairs;
  • triangular FX or crypto routes.

A futures-cash relationship can be written as:

\[F = S e^{(r-q)T}\]

where \(F\) is futures price, \(S\) is spot price, \(r\) is financing rate, \(q\) is dividend or carry yield, and \(T\) is time to maturity.

A triangular arbitrage loop exists when:

\[R_{A \to B} \cdot R_{B \to C} \cdot R_{C \to A} > 1 + \text{fees} + \text{slippage}\]

In efficient markets, these opportunities are tiny and short-lived. The strategy is less about discovering the formula and more about latency, inventory, execution, and operational reliability.

11. Illegal and manipulative strategies

Not all fast trading is legitimate.

Spoofing means placing orders with intent to cancel before execution to create a false impression of supply or demand. U.S. law explicitly targets spoofing under the Commodity Exchange Act’s anti-spoofing provision. (法律資訊學院)

Layering is a spoofing variant using multiple price levels.

Momentum ignition attempts to trigger other algorithms into chasing a move.

Quote stuffing floods the market with excessive messages, potentially degrading competitors’ ability to process the book.

A useful boundary is:

\[\text{Legitimate quoting} = \text{bona fide willingness to trade}\] \[\text{Spoofing} = \text{order placement with intent to cancel and mislead}\]

The hard part in enforcement is proving intent.

12. Flash Crash and fair-weather liquidity

The May 6, 2010 Flash Crash showed that electronic liquidity can disappear or recycle rapidly under stress. The CFTC-SEC joint report describes a large automated sell program in E-mini S&P 500 futures and a sharp cross-market liquidity event. (證券交易委員會)

Kirilenko, Kyle, Samadi, and Tuzun later studied the event and concluded that HFTs did not cause the Flash Crash but contributed to the rapid transmission of selling pressure by aggressively trading and passing inventory around. (Wiley Online Library)

The lesson is not “HFT always removes liquidity.” The lesson is that HFT liquidity is often conditional. A modern market maker is not a traditional specialist with a broad obligation to stabilize the market. It can reduce size, widen quotes, or stop quoting when toxicity and volatility rise.

13. Regulation: Reg NMS, tick sizes, access fees

U.S. equity market structure is deeply shaped by Regulation NMS, including the NBBO, protected quotations, access fees, and routing incentives.

The SEC adopted 2024 amendments to Regulation NMS that reduce access-fee caps for protected quotations in NMS stocks priced at or above \(\$1.00\) to \(\$0.001\) per share. (證券交易委員會)

As of June 11, 2026, the SEC stated that it extended temporary exemptive relief from compliance dates for Rules 600, 610(c), and 612 until the first business day of November 2027, and also proposed rescinding Rule 611, the trade-through rule. (證券交易委員會)

This matters for HFT because tick size and fee caps alter:

  • queue value;
  • maker-taker economics;
  • spread capture;
  • routing incentives;
  • the profitability of rebate strategies;
  • the cost of crossing the spread.

A simple fee-adjusted spread is:

\[\text{Net spread} = (P_{\text{ask}} - P_{\text{bid}}) + \text{maker rebates} - \text{taker fees} - \text{clearing and routing costs}\]

Small changes in fees can flip the profitability of a high-turnover strategy.

14. Speed bumps and frequent batch auctions

One policy response is to slow specific parts of the market. IEX was approved as a national securities exchange in 2016, and its model includes a 350-microsecond delay designed to reduce stale-quote latency arbitrage. (聯邦公報)

Another response is to change time itself: frequent batch auctions. In the Budish-Cramton-Shim design, orders are accumulated and cleared at frequent intervals by uniform-price auction. The central idea is:

\[\text{Continuous time} \Rightarrow \text{race on speed}\] \[\text{Discrete batch time} \Rightarrow \text{competition on price}\]

The tradeoff is market-design complexity. A batch auction can reduce the value of nanosecond races, but it also changes execution priority, routing logic, and the behavior of liquidity providers.

15. Crypto HFT and MEV

Crypto reproduces many traditional microstructure problems but adds a new layer: block ordering.

Centralized crypto exchanges look like fragmented electronic venues: order books, maker-taker fees, API latency, cross-exchange arbitrage, perpetual futures, and funding-rate trades.

Decentralized exchanges are different. Uniswap, for example, is an automated market maker rather than a traditional order book. Its documentation describes AMMs as smart contracts that let users swap tokens and provide liquidity directly on-chain. (Uniswap Developers)

The classic constant-product AMM formula is:

\[x y = k\]

where \(x\) and \(y\) are token reserves and \(k\) is constant before fees and liquidity changes.

MEV, or maximal extractable value, is the value block producers or builders can extract by ordering, including, excluding, or inserting transactions. Ethereum documentation notes that sandwich trading worsens user execution through increased slippage, while proposer-builder separation relates MEV to transaction ordering and validator profitability. (ethereum.org)

Flashbots describes itself as an organization formed to mitigate negative externalities from MEV, and MEV-Boost is described as an initial implementation of proposer-builder separation for proof-of-stake Ethereum. (Flashbots)

In TradFi, the race is often:

\[\text{Who reaches the matching engine first?}\]

In DeFi, the race can become:

\[\text{Who controls or influences transaction ordering inside the block?}\]

That makes MEV an on-chain cousin of latency arbitrage.

16. Does HFT create value?

The best answer is conditional.

HFT creates value when it:

  • narrows spreads;
  • improves price discovery;
  • supplies resilient liquidity;
  • reduces execution costs;
  • arbitrages fragmented prices into alignment.

HFT destroys or transfers value when it:

  • invests heavily in socially redundant speed races;
  • snipes stale quotes without improving fundamental price discovery;
  • increases adverse selection for slower liquidity providers;
  • creates fragile liquidity that disappears under stress;
  • manipulates the book through spoofing or layering.

The empirical literature supports the idea that low-latency and algorithmic trading can improve normal-time market quality, but the market-design literature shows that continuous-time trading can still create structural speed races. (Wiley Online Library)

So the correct conclusion is not:

HFT is good.

or:

HFT is bad.

It is:

HFT market-making and HFT latency arbitrage are economically different activities that happen to use similar technology.

17. Practical learning path

For a trader or researcher, learn HFT in this order:

  1. Limit order book mechanics.
  2. Order types and matching priority.
  3. Bid-ask spread, market impact, and adverse selection.
  4. Queue position and fill probability.
  5. Avellaneda–Stoikov market-making.
  6. Order-book imbalance and microprice.
  7. Toxicity and VPIN-style flow metrics.
  8. Cross-asset and cross-venue arbitrage.
  9. Reg NMS, tick size, maker-taker fees, and routing.
  10. Flash Crash, circuit breakers, speed bumps, and frequent batch auctions.
  11. Crypto market structure and MEV.

The hard part is not writing the formula. The hard part is estimating whether the formula survives:

\[\text{fees} + \text{slippage} + \text{latency} + \text{queue loss} + \text{adverse selection} + \text{operational risk}\]

A strategy that looks profitable before these terms is usually not a strategy. It is a backtest artifact.

18. Final takeaway

HFT is market microstructure under a microscope. At human speed, markets look like prices moving. At HFT speed, markets look like queues, timestamps, cancellations, stale quotes, rebates, matching rules, and adverse-selection games.

The central equation is:

\[\text{Edge} = \text{information} + \text{speed} + \text{queue priority} + \text{risk control} - \text{costs} - \text{adverse selection}\]

If the edge comes from tighter spreads and better price discovery, HFT can improve markets. If the edge comes only from racing to pick off stale quotes, the social value is much weaker. If the edge comes from misleading the book, it is not just weak value — it may be illegal.