AI in Trading: Where It's Mature, Where It's Hype

AI in trading isn’t new — quantitative hedge funds have used machine learning for two decades. The 2026 hype wave (LLM-driven traders, autonomous market-making, AI portfolio managers) is mostly an extension of techniques already in production. Some genuinely new things; a lot of repackaging.

The honest view.

Where AI in trading is mature#

Execution algorithms. TWAP, VWAP, smart order routing, dark-pool routing — all use ML to optimize against market conditions. Decades of deployment. Marginal improvements continue.

Quantitative strategies in liquid markets. Statistical arbitrage, factor models, momentum and mean-reversion strategies with ML feature engineering. Routine at quant funds.

Risk modeling. VaR estimation, factor decomposition, stress testing. Production-credible ML applications.

Surveillance and compliance. Pattern detection in trading activity for market manipulation, insider trading, regulatory compliance. Established techniques with steady improvement.

Where the 2026 hype lives#

LLM-driven discretionary trading. Vendors pitching “GPT trader” outcomes. The evidence for sustained outperformance is weak. LLMs help analyze documents and summarize; they don’t have edge in price prediction.

Fully autonomous market-making. Some research; not yet a credible production deployment outside high-frequency niches.

AI-driven portfolio management for retail. The “robo-advisor with AI” pitch. Mostly marketing on top of standard portfolio construction.

AI signal generation from social media / news. Some real signal; mostly arbitraged away within months of becoming public.

What’s genuinely new#

Foundation-model-driven research. Using LLMs to read earnings calls, 10-Ks, sell-side research at scale and extract structured signals. Real productivity multiplier for human research teams.

Multimodal models for alternative data. Satellite imagery, transcript analysis, supply-chain monitoring. Real edge for funds that build the data pipelines.

Reinforcement learning for execution. Marginal but real improvements on traditional algos.

Generative AI for synthetic training data. Useful for low-frequency strategy backtesting.

The infrastructure under credible deployments#

The math, not the model, is most of the work:

• Tick data infrastructure (terabytes per day for active strategies)
• Sub-microsecond execution paths for HFT-adjacent work
• Backtesting infrastructure that prevents look-ahead bias
• Risk infrastructure that runs in real-time
• Compliance and surveillance integrated with the trading desk

Our data engineering practice builds parts of this for institutional clients.

Where compliance matters#

Trading AI sits in heavily regulated territory:

• MiFID II (EU) requires algorithmic trading governance
• SEC Reg AT (proposed; partial implementation) for US algorithmic trading
• FINRA rules on supervisor responsibility for algos
• AML/MAR market-abuse surveillance

Models that operate outside this framework produce regulatory risk. The compliance discipline is part of the architecture.

What we ship for institutional clients#

For trading-AI engagements:

• Data infrastructure for tick, news, alternative data
• Execution-algorithm integration with OMS/EMS
• Backtesting infrastructure with bias prevention
• Risk-monitoring pipelines
• Surveillance integration for compliance
• Documentation supporting model risk management

Where retail-AI investment products fit#

The retail-investor “AI portfolio” category is largely marketing. The underlying products are:

• Diversified ETF portfolios with rule-based rebalancing
• Tax-loss harvesting (a useful service)
• Risk-tolerance-driven asset allocation

These are good products. They’re not “AI” in any meaningful sense beyond marketing. Customers should evaluate them on standard portfolio-construction merits.

The honest summary#

Trading AI is real, productive, and has been for years. The institutional users with the infrastructure see real edge. The retail products labeled “AI” are mostly standard portfolio management. The breakthrough claims should be evaluated against the well-established baseline of quantitative trading — most “AI trader” pitches don’t survive that comparison.

---

Trading AI works at institutional scale with the right infrastructure. Retail “AI trader” pitches mostly don’t. Our team builds infrastructure for institutional trading and risk programs. Tell us about the program.