How Hudson River Trading Leverages AI for Market Making and Prediction

Hudson River Trading's Business Model

• 🎯 Hudson River Trading operates as a quantitative, automated proprietary trading firm, primarily functioning as a market maker.
• 💡 They provide a utility service by being ready to buy or sell any financial instrument (stocks, futures, options, crypto, bonds) at tight, competitive prices, profiting from the bid-ask spread.
• 🚀 This is likened to Amazon's role as a sophisticated middleman, facilitating the movement of assets through time and space.

Evolution from Traditional Algorithmic Trading to AI

• 🧠 Historically, trading relied on handcrafted features based on human intuition and simple mathematical models like linear regression.
• ✨ Hudson River Trading began integrating AI and machine learning in 2013-2014, moving towards models that consume vast amounts of data without human bias.
• 📈 Over time, this AI-driven approach has entirely overtaken traditional methods, with models now consuming all available data to drive trading decisions.
• 🤖 This process is compared to how ChatGPT is trained, by consuming vast internet data to produce emergent capabilities.

AI's Role in Prediction and Execution

• 📊 AI in trading contributes to both execution efficiency and the ability to make sophisticated short-term price predictions.
• ⚡ The models are capable of processing internet-scale datasets from low-level market events, far beyond human capacity for feature engineering.
• 📈 While models are not easily interpretable, they achieve slight statistical edges (e.g., 50.1% accuracy) which, when applied at scale, significantly boost profits.
• ⚠️ For intraday predictions, market data is the most crucial input, reflecting the real-time intentions of buyers and sellers.

Data Sources and Constraints

• 🌐 For intraday trading, market data feeds from exchanges are the primary and most useful data source, not alternative data like news or social media.
• 🗓️ For longer time horizons (days), alternative data such as SEC filings, news feeds, and broker reports become more relevant, though this is outside the speaker's direct expertise.
• ⚡ A major long-term constraint for scaling AI infrastructure is electricity availability, impacting the ability to build and power new data centers.
• 💡 While GPU availability was a crunch in late 2023, supply has ramped up, making electricity a more significant bottleneck.

Competitive Advantage and Risk Management

• 🏆 Competitive advantage now stems from optimizing the entire technology stack, from data collection and storage to model training and serving, rather than just low latency.
• 🤝 Talent is a key competitive factor, requiring individuals who are both strong researchers and engineers.
• 🛡️ Robust risk management and layered defenses are critical to prevent catastrophic failures, drawing lessons from historical incidents like Knight Capital's rogue algorithm.
• ⚖️ Regulatory compliance is paramount, with a deep respect for rules and a low tolerance for operational errors that could jeopardize trust and market access.

The Future of AI in Trading

• ⏳ The distinction between AI research and engineering is blurring, with research ideas intimately connected to implementation.
• 🤖 While LLMs are powerful, their latency can be too high for real-time market predictions, and their training on historical data poses challenges for reliable backtesting.
• 🧑‍💻 Despite advancements, humans remain involved in high-speed trading, particularly in niche products, highlighting the complexity of fully automating all trading decisions.
• 📈 Hudson River Trading has expanded beyond purely high-frequency trading to include medium-frequency trading, using shorter-term models to inform longer-term strategies.