AI and Quant Trading: A Revolution, Measured

01 · Executive summary

The automation revolution in trading already happened, largely between 1999 and 2015, and the market has been machine-intermediated ever since. High-frequency trading was estimated at roughly 55% of US equity volume by SEC research staff a decade ago. Fifty-nine per cent of the 9.6 trillion dollars traded daily in foreign exchange executes electronically (BIS Triennial Survey, April 2025). Principal trading firms carry more than half of the electronic interdealer Treasury market. What is happening now is a second, different revolution: machine learning and large language models are migrating from execution, where machines already rule, into research, signal generation and the middle office.

Our central conclusion is deliberately double-edged. AI is transforming the cost structure and information metabolism of trading, and it is not repealing the arithmetic of alpha. Documented evidence shows extraordinary machine-driven profitability at capacity-constrained scale (Medallion's documented 66% average gross returns on a fund deliberately capped near 10 billion dollars) alongside equally documented mediocrity when the same intelligence is scaled (the same firm's public funds losing 19 to 31% in 2020; the flagship AI-powered ETF underperforming the S&P 500 by roughly 4 to 5 percentage points a year since 2017; ESMA finding no statistically significant alpha in EU funds that market AI use). The revolution is real. It is a revolution in speed, cost, breadth and market share, concentrated in a shrinking number of firms, and not a revolution in aggregate excess return, which remains bounded by capacity, crowding and signal decay. By 2040 we expect trading to be almost fully machine-executed, substantially machine-researched, and still human-owned at the level of risk appetite, because the scarce resource in markets is not intelligence; it is uncorrelated information and the right to be wrong in size. Postulated

02 · The baseline: markets are already machines

ObservedAny claim that AI will "automate trading" must first confront how automated trading already is. The official record, by asset class:

The physical infrastructure of this baseline is itself instructive. The latency arms race has been quantified at the microsecond level: races lasting 5 to 10 microseconds occur roughly once a minute per FTSE-100 stock, the fastest six firms win more than 80% of them, and latency arbitrage taxes global equity investors by an estimated 5 billion dollars a year (Aquilina, Budish and O'Neill, Quarterly Journal of Economics, 2022). A market that spent 300 million dollars on a straighter fibre route between Chicago and New Jersey does not need AI to become fast. It needs AI, if at all, to become smarter about what to do at speed.

03 · What the machines have earned

ObservedThree data clusters define the ceiling, the middle and the floor of machine intelligence in markets.

The ceiling: capacity-constrained brilliance

Renaissance Technologies' Medallion fund, as documented by Zuckerman (2019), averaged roughly 66% gross and 39% net annually from 1988 to 2018, behind fees of 5 and 44. The critical, under-quoted fact is the denominator: Medallion is deliberately capped near 10 billion dollars, distributes profits annually, and is closed to outsiders. In 2020 Medallion returned 76% while Renaissance's scalable, outside-facing funds lost 19 to 31%, prompting redemptions approaching 15 billion dollars. One firm, one year, both truths: machine alpha exists, and it does not scale.

The middle: the market-making dynasty

The clearest AI-era winners are not funds but machine market-makers, whose edge compounds through flow and infrastructure rather than forecasting. Jane Street's net trading revenue reached 20.5 billion dollars in 2024 and a record 39.6 billion in 2025, reported as exceeding JPMorgan's trading haul. Citadel Securities earned a record 9.7 billion in 2024 and roughly 12 billion in 2025. These figures, from bond-investor disclosures relayed by Bloomberg and the FT, represent the industrialisation of intermediation: the profit pool has moved from predicting the market to being the market.

The floor: scaled intelligence without edge

The AI Powered Equity ETF, running on machine intelligence since October 2017, has returned roughly 10% annualised against roughly 14 to 15% for the S&P 500, underperformance of 4 to 5 percentage points a year across a full cycle. ESMA's census is broader and harsher: of the EU's 10.8 trillion euro UCITS universe, only 106 funds, 0.1% of assets, formally promoted AI or machine-learning use as of early 2024, and those funds showed no statistically significant outperformance against peers (ESMA, February 2025).

04 · Why intelligence does not equal alpha

ObservedThe gap between the ceiling and the floor has causes, and they are quantified in the academic record. Markets are a hostile environment for machine learning in four specific, structural ways.

• Signals die in public. Published anomaly returns fall roughly 26% out-of-sample and 58% post-publication (McLean and Pontiff, Journal of Finance, 2016). The same decay has now been observed for LLM signals: the profitability of ChatGPT-derived news sentiment strategies declined as adoption spread, consistent with the market pricing the signal away (Lopez-Lira and Tang, 2023 onward). Every alpha democratised is an alpha destroyed.
• Backtests lie by construction. With enough strategy trials, a high in-sample Sharpe ratio is expected even with zero true skill; the probability of backtest overfitting is quantifiable and pervasive (Bailey, Borwein, López de Prado and Zhu, 2014 to 2017). The deflated Sharpe ratio exists precisely because the industry's default evidence, the backtest, is systematically inflated. AI multiplies the number of trials, and therefore multiplies the overfitting surface, faster than it multiplies insight.
• The data regime shifts beneath the model. Financial time series are non-stationary with low signal-to-noise; the relationships a model learns degrade exactly when they matter, at regime breaks. The 2018 to 2020 quant winter is the documented case: Renaissance's public funds down 19 to 31%, a leading factor fund shrinking 92% from its 2018 peak, AQR's assets falling from roughly 226 to 140 billion dollars across the drawdown.
• Capacity is the tax on truth. The strongest documented machine alpha in history is capped at 10 billion dollars by its own operators. That decision, made by the people with the best information about scalability ever assembled, is the single most eloquent statement in quantitative finance.

05 · The execution layer: where AI already trades

ObservedOne tier of trading AI is neither hype nor future: optimised execution. JPMorgan's LOXM, rolled out from 2017, was the first major bank equities algorithm trained with reinforcement learning, and the bank reported execution-cost improvements in internal trials without ever publishing a quantified figure, a silence we respect by citing none. The academic foundation predates the current wave by two decades: reinforcement learning for optimised trade execution demonstrated significant improvement over naive order placement as early as 2006, and a 2025 survey now catalogues a mature field. In foreign exchange, the BIS Markets Committee estimated execution algorithms at 10 to 20% of global spot turnover as far back as 2020. Execution is the ideal habitat for machine learning in markets: the objective is well defined (minimise implementation cost against a benchmark), the feedback loop is fast and labelled, the horizon is minutes rather than regimes, and the adversary is market microstructure rather than the future. Every property that makes execution tractable is a property alpha generation lacks, which is why the execution layer industrialised quietly while the forecasting layer keeps disappointing. The distinction between the two layers, routinely blurred in commentary, is the difference between a solved engineering problem and an open scientific one.

06 · The second wave: LLMs on the desk

ObservedThe generative wave is entering through research and workflow, not the order book. AIMA's September 2025 survey found 95% of 150 fund managers, with roughly 788 billion dollars of combined assets, using generative AI in their work, up from 86% in 2023, with 58% expecting front-office investment use within a year. Yet ESMA's February 2026 census found that of 847 reported AI use cases among EU firms, only about ten touch algorithmic trading and three touch high-frequency trading. Adoption is near-universal; autonomy remains near-zero. The academic frontier explains the appeal: GPT-4-class models predicted the direction of earnings changes at roughly 60% accuracy against 53% for human analysts in backtest (Chicago Booth working paper, 2024), and LLM news-sentiment signals showed statistically significant predictive power before crowding eroded them. The models read faster than any analyst ever will. Whether reading faster remains valuable once every desk reads at the same speed is the crowding question this paper keeps returning to.

07 · The stability file

ObservedAutomation has a documented instability signature, and supervisors expect AI to sharpen it. On 6 May 2010 a single automated sell programme in E-mini futures, executing 75,000 contracts by volume-participation logic, helped drive the Dow down roughly 9% intraday, with over 20,000 trades later broken at prices more than 60% from pre-crash levels (SEC and CFTC final report, 2010). On 15 October 2014 the 10-year Treasury yield fell 16 basis points and rebounded within minutes, with no identifiable catalyst, in a market where principal trading firms carried the majority of electronic volume (Joint Staff Report, 2015). The regulatory response built after 2010, limit-up-limit-down bands, market-access risk checks, kill switches and order-to-trade limits under MiFID II, was designed for fast machines executing human strategies.

The defensive architecture deserves specification, because its adequacy is the open question. In the United States: the Market Access Rule (2010) mandates pre-trade risk checks; limit-up-limit-down price bands and revised market-wide circuit breakers (2012) constrain single-name and index moves; Regulation SCI (2014) imposes systems-integrity duties on exchanges and large venues; and the Consolidated Audit Trail (phased from 2020) gives supervisors a full order-level record after the fact. In the European Union, MiFID II's Article 17 and its technical standards require algorithm testing, kill switches, high-frequency registration and venue-level order-to-trade limits. Every element of this stack was designed against the failure mode of 2010: fast machines executing badly specified human instructions. None of it was designed against the failure mode the supervisors now describe, many institutions independently reaching the same decision at the same moment because their models share a prior. Circuit breakers halt cascades; they do not decorrelate them. Observed

The new concern, stated in near-identical language by the IMF (October 2024), the FSB (November 2024), the ECB (May 2024) and Federal Reserve staff (2025), is correlation: many firms running similar models, trained on similar data, supplied by a handful of vendors, reacting identically in stress. The IMF recommends recalibrated circuit breakers and better monitoring of nonbanks precisely because AI-heavy intermediation is migrating toward hedge funds and principal trading firms outside the bank supervisory perimeter. The IMF also documents the direction of travel: the AI share of algorithmic-trading-related patent filings rose from roughly a fifth in 2017 to more than half in every year since 2020.

08 · Market psychology and the sentiment regime

Prices are made by participants whose cognition is bounded and whose incentives are structured, and nowhere is that more consequential than in a market pricing its own automation. Three psychological structures deserve explicit statement, because each is documented rather than assumed.

• The AI label sells what the AI record does not support. ESMA's census found funds marketing AI delivered no measurable outperformance, and the SEC's first AI-washing enforcement (two advisers fined a combined 400,000 dollars in March 2024 for overstating their AI use) confirms that the label itself has commercial value detached from capability. When a technology's name is worth more than its measured contribution, the market is in a narrative regime with respect to that technology. Behavioural diagnosis: availability heuristic and herding, with the AI story supplying the availability.
• Survivorship shapes the entire public conversation about machine alpha. The visible record is Medallion and Jane Street; the invisible record is the majority of quant launches that closed, the factor complex's 92% peak-to-trough shrinkage in one documented case, and the redemption cycle that followed 2020. Allocators anchoring on the visible tail are making a base-rate error the academic literature has quantified for them: most backtested edges do not survive contact with capital.
• Reflexivity runs through the machines themselves. When many models learn from the same data and act on the same signals, price action increasingly reflects the models' shared prior rather than external information, which is precisely the mechanism behind the supervisors' herding concern. The 2018 to 2020 quant winter can be read as an early instance: crowded factor positioning unwinding through the very models that built it. AI deepens the loop because foundation models share training lineages in a way hand-built factor models never did.

PostulatedSentiment-adjusted judgment: enthusiasm for AI-in-markets narratives is in a belief regime, while measured autonomous deployment remains minimal (ESMA's ten use cases out of 847). That spread between story and state is itself information. Historically, capability gets overpriced early and infrastructure gets underpriced early; the trading-infrastructure winners of Section 3 were being built, largely unremarked, while the previous decade's narrative capital chased forecasting funds that no longer exist.

09 · The counter-thesis, steel-manned

The strongest case that this time is different, stated at full strength: large language models are not another factor model. They ingest unstructured information, filings, calls, news, supply-chain chatter, that priced slowly precisely because it required human reading, and the documented early results (roughly 60% directional accuracy on earnings changes against 53% for analysts; statistically significant news-sentiment predictability) are exactly what an unexploited information channel looks like before capital crowds it. Moreover, the decisive inputs of the next decade, proprietary data exhaust and compute scale, are not equally distributed, so the crowding that killed public anomalies may not operate: an edge fed by private data does not decay through publication. On this view, two or three firms could assemble Medallion-class economics at ten times Medallion's capacity, and the 15% bull scenario below is far too thin.

PostulatedOur response concedes the mechanism and disputes the equilibrium. Private-data moats are real, but the documented history of execution advantage shows moats migrating toward infrastructure owners, who monetise them as intermediation rather than as directional alpha, which is why the profit record of the machine era sits with market-makers. And the LLM information channel, unlike a proprietary dataset, is being distributed to every desk simultaneously through commercial APIs; the Lopez-Lira decay evidence suggests the crowding clock started the day the papers published. We hold the bull case at 15% deliberately: high enough to respect the mechanism, low enough to respect the graveyard.

10 · Scenarios to 2040

PostulatedProbabilities are EWC judgment, basis: the deployment and stability record above, base rates from three decades of market automation, and supervisor trajectory. They sum to 100%. Milestone phasing follows the EWC convention: 2026 to 2030, 2030 to 2035, 2035 to 2040.

Path dependency note: the bear scenario, if it occurs early (before 2030), likely produces the constraint regime that prevents the bull scenario from ever being tested. The order of events matters more than their probabilities.

11 · Conclusions

First, the honest headline: the machines won the execution war years ago, and the current AI wave is about cognition, not speed. Its measured beachhead is research productivity, where the evidence of skill is real, and its measured limit is crowding, which converts every shared signal into no signal.

Second, the profit pool tells the structural truth. The decade's great trading fortunes accrued to machine market-makers, not machine forecasters. Flow, infrastructure and cost curves compound; predictions decay. We expect the 2040 hierarchy to preserve that ordering. Postulated

Third, the stability question is now the main event. Every major supervisor has converged on the same scenario, correlated models deleveraging together, and the market's plumbing was built for a different failure mode. Whether the 2030s deliver the industrialised-intelligence base case or the correlated unwind depends substantially on whether circuit-breaker and model-governance reform arrives before the first AI-native stress event. That race, between governance and correlation, is the single most consequential variable in market structure today.

EWC Investments Think Tank publishes independent research for educational purposes. EWC does not manage client capital, does not operate an investment fund, and does not provide regulated portfolio management or personalised investment advice. Forward-looking statements are inherently uncertain; probabilities stated are analytical judgments, not guarantees. Private-firm figures cited are press-documented, not audited public filings. Figures carry their as-of dates and may have changed. © EWC Investments, 2026.