Machine learning has taken over Wall Street. Here's how the latest models are reshaping market structure.
Wall Street has quietly finished a transition that started decades ago with simple program trading: according to SEC data, AI-driven algorithms now account for roughly 73% of US equity trading volume, up from 60% in 2023, and the systems doing that trading bear little resemblance to the rule-based algorithms that dominated the previous decade.
From rule-based signals to language-model reasoning
The older generation of algorithmic trading ran on relatively simple quantitative signals: price momentum, volume patterns, order-book imbalances, executed at speeds no human could match but reasoning in ways that were fundamentally mechanical. The current generation layers large language models on top of those same quantitative signals, feeding in earnings-call transcripts, breaking news, social-media sentiment, and even satellite imagery of parking lots and shipping ports. Firms like Citadel, Two Sigma, and Renaissance Technologies are effectively running continuous, automated equity research at a volume and speed no team of human analysts could replicate, synthesizing unstructured information into trading signals within seconds of it becoming public.
The performance gap is now measurable
This isn't a marginal edge. A Bloomberg analysis of 50 AI-focused hedge funds found they outperformed the S&P 500 by an average of 4.2 percentage points in 2025, and notably did so with significantly shallower drawdowns during market corrections. The mechanism behind the smaller drawdowns is instructive: because these systems process new information faster and more comprehensively than human traders, they can reposition ahead of a broader market reaction rather than scrambling to catch up to one, which matters most precisely during the volatile periods when human decision-making tends to degrade under pressure.
Regulators are watching the correlation risk
The SEC's response has moved from monitoring to active rulemaking. Proposed rules would require AI-powered trading systems to undergo stress testing similar to bank capital requirements, aimed at ensuring predictable behavior during market crises rather than just profitable behavior during calm ones. The specific worry animating this push is correlated failure: if a large share of trading volume is driven by models trained on similar data and architectures, a shared blind spot or a simultaneous reaction to the same signal could amplify a market move well beyond what any single firm intended, a systemic risk that didn't exist when trading strategies were more heterogeneous.
What this means for the gap between institutions and individuals
The tools powering institutional trading desks are also becoming available, in narrower form, to individual researchers and smaller funds. Vincony's Sentiment Analyzer provides real-time sentiment analysis across news, social media, and earnings calls, the same category of unstructured-data processing that underpins hedge fund trading models, now accessible without a Two Sigma-sized data science team behind it. That doesn't close the gap entirely, institutional desks still have proprietary datasets, colocated servers, and lower-latency infrastructure that individual researchers simply cannot replicate, but it meaningfully narrows the information-processing advantage that used to belong exclusively to the largest players.
The infrastructure arms race behind the scenes
None of this performance comes cheap to produce. Running LLM-based sentiment extraction across every earnings call, news wire, and social media platform in real time, then feeding that into a trading decision within milliseconds, requires infrastructure investment that only the largest funds can currently justify. This is quietly reshaping the competitive landscape of quantitative finance: the barrier to entry is shifting away from having the best statistical model, which is increasingly commoditized, and toward having the fastest, most comprehensive data pipeline feeding that model, which is not.
An open question about market structure itself
As the gap between institutional and retail capability narrows, a harder question emerges about what market efficiency even means when a large majority of volume is AI-driven and increasingly influenced by tools available to retail participants as well. Whether this makes markets fairer, by spreading analytical capability more broadly, or simply faster and more correlated in ways that create new forms of systemic fragility, is likely to be one of the defining market-structure debates of the next several years, and the SEC's stress-testing proposals suggest regulators are already betting on the latter.