Is HFT still profitable in 2026?

As an industry yes, but margins have compressed for over a decade. Pure speed in mature equities is a commoditised utility; the live edge sits in newer venues and smarter microstructure signals, not raw latency. The honest, dated answer.

The idea

Is HFT dead? The short answer

No, but the question is malformed. "HFT" bundles several different businesses, and they are in different states. Some segments are effectively closed to newcomers (pure speed); some are alive but consolidated into thin-margin utilities (market making); some are crowded (stat arb); and some are genuine open frontiers (crypto, prediction markets). Asking "is HFT profitable" without naming the segment cannot be answered.

Intuition first: "is HFT still profitable" is like asking "is software still profitable": for whom, doing what, where? The honest answer requires splitting the field into the things people actually mean, and that split is the entire structure of this page and of the strategies topic. What is true across all segments: per-trade margins have compressed steadily over the decade as competition, technology and regulation matured; the field has consolidated; and a generic "I'll do HFT" has no edge. Every surviving edge is specific to a segment, a venue and an infrastructure. That consolidation-and-compression story is real and important, and it is why most "HFT is dead" takes have a grain of truth wrapped around a wrong conclusion.

So the headline of this page is: HFT is not dead, but "HFT" is not one thing. Profitability has bifurcated. The rest is the segment-by-segment ledger: which edges are dead, which are commoditised utilities, which are still alpha, and where the open frontier is.

Pure speed (latency races): consolidated, brutally hard

Commoditised to a near-utility, effectively closed to newcomers. Pure latency arbitrage and speed-based queue racing on lit venues is a decade-deep infrastructure arms race won by a handful of firms with colocation, FPGA hardware and inter-datacentre microwave links. The edge is real, but it is an infrastructure edge you out-spend, not out-think, and a newcomer cannot.

The mechanism is winner-take-all in microseconds. Beating the field by 2 microseconds wins nearly every contested fill; losing by 2 microseconds wins none. So the game collapsed to "who has the fastest path", and the fastest path costs tens of millions in colocation and FPGA infrastructure and microwave links. By 2026 a small set of firms (see market participants) have spent a decade and a fortune on the last microseconds, and the marginal gain from the next microsecond is shrinking. The arms race has reached diminishing returns even for the incumbents. For a newcomer the entry barrier is total. This segment is the strongest case for "the easy money is gone".

Where any speed edge survives: briefly, on new fragmented venues before the incumbents arrive (some crypto venues), and as a defensive necessity for market makers, where you need to be fast enough not to be picked off, even if speed is not your edge. The deeper economics are in the economics of HFT.

Market making: alive, but industrialised

Structurally alive, but industrialised into a low-margin utility on lit venues. Someone must always quote, so market making does not "die", but on lit equities and futures it is a scale-and-latency business run by a few firms on razor-thin per-trade margins. The genuine open frontier for an independent is making markets in crypto and prediction markets.

Why it is structural: every order-driven venue needs liquidity on both sides, and the spread is real compensation for inventory risk and adverse selection. That trilemma is invariant: it cannot be competed away the way a single signal can. So market making is permanent terrain, not a decaying edge. Why it is industrialised on lit venues: the per-trade margin is tiny, so the business is a law-of-large-numbers game: vast trade counts across thousands of names on fast infrastructure. That favours scale and capital, so lit-equity market making consolidated into a utility (see equities & futures). The formula is textbook (Avellaneda–Stoikov); the edge is execution, speed and scale.

Where it is still open: crypto and prediction markets, where the colocation, prime-broker and licensed-data barriers are absent or cheap, books are thinner, and competition is shallower. This is the single most important "still profitable" claim on the page for the independent reader: the maths is the same, the open venue is the difference. See the markets hub.

Order-flow / information market making: the live alpha

The most genuinely alpha-bearing segment in 2026. Reading order-flow toxicity and imbalance to quote around a better fair value is a renewable edge: the flow keeps coming, and interpreting it well is a research problem, not a pure speed problem. This is where the modern market-making edge actually sits, and where AI helps most.

The mechanism: instead of quoting symmetrically around the mid, you quote around a fair value that reads the book and the flow (the microprice, order-flow imbalance, PIN/VPIN) so you are adversely selected less. This is Market Making II, the information family. Why it is alive: the edge is renewable. Order flow arrives continuously and carries information continuously, so a better read of it is a durable advantage, unlike a one-off signal that is arbitraged and gone. And it is a research edge (better fair-value and toxicity estimation), which is exactly where machine learning and AI add genuine value: better inputs to the quote, off the latency-critical hot path.

The caveat: it still requires clean L3 data and low-enough latency not to be picked off, so it is not barrier-free, but the barrier is research and data quality, not tens of millions in microwave links. The most accessible version is on the open venues (crypto, prediction markets), where the data is gettable.

Statistical arbitrage: crowded, edge migrated

Crowded and largely commoditised in liquid equities; alive at harder horizons and venues. The classic pairs trade has been public since the 1990s and is arbitraged in milliseconds in liquid names. The surviving stat-arb edge has migrated to faster horizons, crypto, ETF and index mechanics, and harder-to-arb corners where a better signal still pays.

Why the classic trade is dead: cointegration-based pairs trading is in every textbook, so the obvious relationships are crowded and the spreads close before a newcomer can act. The edge is not in knowing the technique (everyone does) but in finding relationships others have not, faster, on cleaner data. (The IX-COINT explorer lets you feel how thin the surviving edge is once costs bite.) Where it survives in 2026: cross-exchange and cross-pair mean reversion in crypto (the classic trade, on a less-arbitraged venue), ETF and index rebalancing mechanics, and very-short-horizon intraday reversion that is mechanical and cost-sensitive.

The edge migrated toward the harder, less-crowded corners. That is the general pattern of alpha decay, and it applies to every segment on this page, not just stat arb. The technique is permanent; any specific spread's edge is on a clock.

Event and news trading: live but harder

Live but harder. The slow-repricing edge is gone, the interpretation edge remains. Simple drift after scheduled events is competed away; the surviving event-trading edge is latency-to-react plus correct interpretation of machine-readable news, increasingly an AI/NLP problem. The cleanest venue is prediction markets, where the contract is the event.

Why the easy version died: the post-announcement drift that textbooks describe was a slow repricing you could ride; it has been arbitraged down. What is left is reacting correctly and fast the instant news lands, and "correctly" is the hard part, because the fast-and-wrong reaction loses. Where it is alive: in the interpretation layer, where machine learning and LLMs genuinely help parse news into a directional view faster than a human; and natively on prediction markets, where the contract is the event and event/news trading is the whole game. See news trading for the mechanics.

Crypto and prediction markets: the open frontier

The genuine open frontier for an independent in 2026. Crypto and prediction markets are where the institutional barriers that wall off lit equities (colocation, prime brokerage, licensed data, large capital) are absent or cheap. The same maths runs on thinner, less-contested books, so a small team can still earn real spread. With real, distinctive risks.

Why these are the answer to "can I still make money in HFT": the barriers are the difference. In crypto you get 24/7 books, gettable data, self-hosted infrastructure and no prime broker; in prediction markets you get an open on-chain order book with bounded payoffs and an underexploited, microstructure-naive field. The competition is shallower than lit equities by a wide margin.

The honest cost: crypto has counterparty risk (your funds sit with the exchange), wash trading and thin, gappy books; prediction markets have bounded payoffs and a unique resolution and settlement risk. These are first-order, not footnotes: the venues are accessible, not safe. But for an independent, this is where the realistic openings are, and it is the practical foundation of the going-independent-in-2026 thesis.

What does it take to be profitable in HFT now?

A specific edge in a specific segment on a specific venue, not a generic "do HFT". You need an edge that is actually alive (per the segments above), the right venue (an open frontier for an independent), gettable clean data, honest backtesting, execution discipline, risk controls, and honest cost accounting. The maths is free; the edge, the venue choice and the discipline are the work.

The decision factorises. Pick a segment that is alive (order-flow market making, market making on open venues, niche stat-arb, event interpretation) times a venue where you can actually compete (crypto, prediction markets) times a real operational stack (data, backtest harness, execution, risk limits, cost accounting). Get any factor wrong and the others cannot save you: a live edge on a closed venue earns nothing, and a great venue with no edge or no risk discipline blows up.

What AI changes: it raises research velocity (the solo-quant enabler) and improves interpretation (news, fair value, toxicity), but it does not change speed (still hardware) or remove overfitting (still kills). See what AI changes for HFT. It moves the operating point; it does not make a dead segment alive. The realistic path for an independent in 2026 is order-flow-aware market making on an open venue, with clean data, a sound inventory model, honest costs and risk discipline. That is exactly the ladder this atlas builds, ending in the datasets and tools you would need to run it. The fuller route map is going independent in 2026.

The margin-compression story, made concrete

The decade's compression, with an illustrative segment comparison (synthetic, directional, orders-of-magnitude, not measured rates).

Pure speed, then and now. A decade ago a latency edge could be measured in milliseconds, and a modest infrastructure budget could win contested fills. By 2026 the contested margin is measured in single-digit microseconds and the winning infrastructure budget is tens of millions, so the number of firms that can profitably play has collapsed from dozens to a handful. The edge did not vanish; it concentrated.

Market-making margin. Per-round-trip gross margin on a liquid large-cap is a couple of basis points (a one-tick half-spread plus a sub-basis-point rebate) so the business only works at enormous trade counts: millions of round trips across thousands of names. A newcomer doing one-thousandth of the volume earns one-thousandth of the (already tiny) margin and bears the same fixed cost: structurally unprofitable at small scale on lit venues. The same model on a thinner crypto book quotes a 20-basis-point spread (an order of magnitude wider per round trip) which is exactly why the open venue is where small scale works.

Stat-arb crowding. A cointegrated pair that yielded a tradeable spread divergence for hours in the 1990s closes in milliseconds in liquid equities in 2026: the half-life of the edge has fallen by orders of magnitude as the field crowded. The surviving version is on venues or horizons where fewer participants compute the same signal (crypto, very-short-horizon, niche corners). This is alpha decay over the decade, made literal. All figures here are illustrative, directional and synthetic (orders of magnitude, not measured rates); real margins, latencies, edge half-lives and costs vary enormously by segment, venue and instrument, and change over time. Reverify against primary sources before relying on any of it. Educational only, not investment advice; no P&L is promised.