By Alex Evangeli
Successful decision making in fixed income markets requires drawing conclusions from fragmented information. This challenge makes the asset class a natural candidate for both machine learning and, more recently, generative artificial intelligence (generative AI). Drawing on my experience building and leading fixed income trading businesses across European and US markets, and on the study of artificial intelligence at Columbia University with Professor Daniel Guetta, I explore where each technology fits, how they can work together, and how risks can be managed.
Machine learning
Machine learning, where algorithms identify patterns in historical data to make predictions, has been applied to fixed income markets for years. One practical application here is bond valuation. Issuers typically have multiple bonds outstanding across different maturities, allowing an issuer specific credit curve to be fitted to the credit spreads of actively traded bonds. This curve provides a benchmark for estimating the fair value of other comparable bonds.
Traditional approaches often rely on interpolation, which can become unreliable when only a small number of bonds trade actively. Machine learning techniques enhance these methods by learning typical issuer curve shapes and pricing relationships from historical data, enabling more accurate fair value estimates especially for illiquid bonds.
Distinguishing meaningful signals from noise is critical. Tree based machine learning methods, such as random forests, may be suited to this problem because they can model complex, nonlinear relationships while also providing measures of feature importance. This helps distinguish valuable predictors from less informative variables.
Boosted tree methods help improve predictive accuracy further, building trees sequentially so that each one corrects the errors of the last. However, even with these complex techniques, combining them with human knowledge remains essential because statistical importance does not always translate directly into economic relevance.
Machine learning therefore remains well suited to structured prediction problems. Generative AI addresses a different challenge altogether.
Generative artificial intelligence
Generative AI helps process and generate insights from large volumes of unstructured information. Large language models (LLMs) produce outputs by learning patterns from vast quantities of training data and generating the most statistically likely next word, one word at a time, based on everything that came before it.
The data fragmentation challenge
Unlike equity markets, fixed income markets contain significantly more individual securities, many of which trade infrequently. Information is dispersed across dealers, data providers, reporting venues and internal systems. In my experience building fixed income trading businesses and supporting technology, the primary challenge has been bringing this fragmented information together quickly to support trading decisions.
Take the construction of primary market ETF creation or redemption baskets. A trader needs to combine real time book positioning, trading signals, and highly specific issuer process requirements into a basket proposal under time constraints. Information is rarely available in a consistent format or via a single source, making standardisation a necessary first step before generative AI can provide meaningful support. Solving this data challenge helps streamline every process that follows.
Once information is properly structured and accessible, an LLM can help support decision making by bringing together insights from multiple sources.
Retrieval augmented generation (RAG) can help solve this problem. Rather than relying solely on what a language model learned during training, RAG can connect the model to proprietary information sources, allowing it to generate responses based upon current data. For fixed income applications, this technology is useful because traders and investment professionals rely on live market information and internally generated signals. When engineered correctly, these technologies are complementary: outputs from machine learning provide one of several data sources a RAG system can draw upon.
For example, a RAG architecture could connect to ETF holdings, index composition and internal basket construction rules, allowing this information to feed downstream into the basket construction process. This reduces manual reconciliation and allows traders to focus on evaluating the output rather than assembling the underlying information. The two approaches work best in sequence. Machine learning answers the question of what a bond is worth today. Generative AI can use that answer together with index constraints, issuer requirements and book positioning to produce a considered output for trader review.
Mitigating the risks
Hallucination is a common problem. LLMs can produce confident, plausible sounding outputs that are factually incorrect. Unlike traditional statistical models, whose assumptions and parameters can often be inspected directly, an LLM’s internal decision making process is difficult to audit. In regulated environments, the lack of auditability presents an obvious compliance issue.
Consistency is a separate challenge, as the same prompt can yield different outputs across runs. This also makes backtesting an LLM based workflow significantly more difficult than backtesting a conventional quantitative model. Trading desks, as a result, are adopting generative AI carefully. In addition, data confidentiality is a significant concern for trading firms, though it falls outside the scope of this article.
FINRA’s 2026 Annual Regulatory Oversight Report addresses ensuring ‘the GenAI solution continues to perform as expected and results in compliant behavior’, suggesting: hallucinations and bias as risks for mitigation and ‘human-in-the-loop review of model outputs’ as part of the monitoring process. Several technical approaches help with the hallucination and consistency problems:
Effective model prompting: allows users to specify outputs using carefully designed prompts, encourages the model to state uncertainty when appropriate, and reduces the likelihood of unsupported answers.
Structured outputs: constraining a language model to respond using a predefined structure rather than unrestricted free text drastically reduces the scope for errors. A validation layer can be applied that halts the process before it reaches the downstream trading workflow.
Tool access: an agentic system gives the model access to purpose built tools such as pricing feeds, bond analytics libraries, and database queries. For example, instead of manually searching multiple systems, a portfolio manager with cash to deploy could request the cheapest healthcare bonds meeting predefined liquidity, duration and amount outstanding constraints. The language model determines which tools to call, while pricing, analytics and portfolio queries are performed by specialist systems.
Validation layers: a second model or rule based system can review outputs before they reach a trading environment, flagging information that falls outside expected ranges or contradicts known data.
Model selection: a tiered approach can be highly effective. Use frontier models for analysing complex events that influence pricing, and smaller, faster models for lower stakes document classification.
Human review: this remains the most reliable check when stakes are high. The human role in fixed income trading has changed significantly in recent years and, in my view, will continue to change given the emergence of this new technology.
Throughout my career, successive waves of technology have automated repetitive tasks while increasing the value of human judgement in more complex situations. I expect generative AI to follow the same pattern. Rather than replacing experienced traders, it is more likely to change where they spend their time, shifting effort from gathering information towards evaluating exceptions and making higher value decisions.
The path forward
Firms that gain the greatest advantage from these new technologies will be those that combine structured data, thoughtful systems design and human validation into workflows. The key is understanding where these new technologies add value, where their limitations lie, and how to integrate them appropriately into trading businesses.
Disclaimer: Views are my own. This is for informational purposes only and should not be relied upon to make investment decisions.
Biography: Alex Evangeli, an experienced voice in ETF market making, has traded the product since 2007. Alex founded and led the fixed income trading business in Europe at Virtu Financial before relocating to New York to trade and lead the development of fixed income trading technology for the firm’s ETF block business. Previously, he held senior positions at Brownstone Investment Group and Cantor Fitzgerald where he was deeply involved in fixed income ETF and cash bond risk trading.
©Markets Media Europe 2026
