Alex Evangeli argues that the fragmented nature of fixed income markets makes them ideal candidates for machine learning and generative artificial intelligence (generative AI). Drawing on his extensive experience in building and leading fixed income trading businesses and studying AI, he explores how these technologies can be applied, how they interact, and how their inherent risks can be managed to enhance trading decisions.
Introduction
Successful decision-making in fixed income markets, characterized by fragmented information, naturally lends itself to machine learning and generative AI. This commentary, by Alex Evangeli, explores the application, synergy, and risk management of these technologies based on his experience in fixed income trading and AI studies.
Machine learning
Machine learning, which uses algorithms to identify patterns in historical data for predictions, has long been applied to fixed income. A key use is bond valuation, where it enhances traditional methods by learning typical issuer curve shapes and pricing relationships, leading to more accurate fair value estimates, particularly for illiquid bonds. Techniques like random forests and boosted trees model complex, nonlinear relationships and indicate feature importance, but human expertise remains critical to ensure economic relevance.
Generative artificial intelligence
Generative AI excels at processing and generating insights from vast amounts of unstructured information. Large Language Models (LLMs) achieve this by learning patterns from extensive training data and predicting the most statistically probable next word, generating text one word at a time.
The data fragmentation challenge
Fixed income markets are complex due to numerous, infrequently traded securities and information dispersed across various sources (dealers, data providers, internal systems). A primary challenge has been the rapid consolidation of this fragmented information to inform trading decisions. Standardizing this data is a crucial first step for generative AI to provide meaningful support. Once structured, Retrieval Augmented Generation (RAG) can connect LLMs to proprietary, live market data and internal signals (including machine learning outputs), streamlining processes like primary market ETF basket construction, reducing manual reconciliation, and allowing traders to focus on evaluating outputs.
Mitigating the risks
Generative AI presents risks such as 'hallucination' (producing confident but factually incorrect outputs) and inconsistency, making auditability and backtesting difficult in regulated environments. Firms are cautious due to these issues and data confidentiality concerns. FINRA suggests 'human-in-the-loop review.' Technical mitigation strategies include effective model prompting, constraining outputs to predefined structures, providing tool access (e.g., pricing feeds, analytics libraries), using validation layers, and employing a tiered model selection approach. Human review remains paramount for high-stakes decisions, shifting traders' roles from information gathering to evaluating exceptions and making high-value judgments.
The path forward
To fully capitalize on these new technologies, firms must integrate structured data, thoughtful system design, and human validation into their workflows. Success depends on understanding the unique value propositions and limitations of these technologies and how to effectively incorporate them into trading operations.
