LLM-GUIDED HEDGE PORTFOLIO CONSTRUCTION IN PREDICTION MARKETS: From Natural Language Risk Descriptions to Implementable Positions
Prediction Markets · LLM Reasoning · Portfolio Construction · Probabilistic Modelling
THE PROBLEM
Many economically important risks have no natural financial hedge.
At the same time, prediction markets increasingly trade contracts on economic, political and social events that may be related to those risks. The difficulty is that these relationships do not initially exist in the mathematical form required for portfolio construction.
A portfolio manager may know that her exposure is affected by interest rates, inflation or political outcomes, while prediction markets trade hundreds or thousands of contracts expressed as natural-language propositions.
The potentially useful hedge may therefore already exist—but be hidden inside a universe organized by language and meaning rather than conventional financial variables.
THE IDEA
The paper treats this as a problem of reconstructing a mathematical space.
An LLM is used not to predict market outcomes directly, but as an elicitation and structural reasoning engine: starting from a natural-language description of the risk, it identifies prediction-market contracts connected to that risk through semantic or causal channels and characterizes the nature of those relationships.
Those judgments cannot simply be inserted into a portfolio optimizer. Individually elicited probabilities and conditional relationships may be mutually inconsistent.
The architecture therefore transforms them into a coherent joint probability model before constructing the covariance structure required for financial optimization. The resulting representation turns a universe of propositions expressed in words into a space on which conventional portfolio mathematics can operate.
FROM MEANING TO MATHEMATICS
The classical minimum-variance hedge is straightforward once the covariance structure between the risk and the available assets is known.
But in this setting, the covariance matrix does not exist until it is constructed. That construction is the substance of the problem.
The framework moves through a sequence:
risk description → semantic and causal structure → coherent probabilities → financial dependence → portfolio optimization → executable positions
The LLM therefore acts at the boundary between two very different representations of information.
On one side is language.
On the other is a mathematical space containing random variables, probabilities, covariances, position sizes and portfolio risk.
The objective is to reconstruct enough structure between them that hidden combinations of prediction-market contracts can become economically useful.
FROM THEORETICAL HEDGE TO IMPLEMENTABLE POSITION
The paper deliberately goes beyond identifying interesting correlations.
Theoretical hedge ratios are translated into executable positions subject to the actual mechanics of prediction-market venues: contract sizes, limits, fees, spreads and market depth.
It also introduces a hedgeability measure: given a particular risk and available contract universe, how much of that risk can the reconstructed market structure theoretically remove?
Importantly, the paper distinguishes this model-implied hedgeability from actual hedge performance. Whether the elicited relationships are correct can ultimately be verified only against realized outcomes.
WHY IT MATTERS
Financial markets traditionally hedge risks for which financial instruments already exist.
Prediction markets suggest a different possibility.
A sufficiently rich universe of event contracts may contain combinations capable of synthesizing exposures that were never explicitly designed as financial hedges.
The challenge is discovering those combinations.
In that sense, the research is another form of pattern extraction: information already exists, but in the wrong coordinate system.
The objective is to transform meaning into mathematical structure—and then search that structure for combinations that conventional markets cannot directly provide.
RESEARCH STATUS
Independent research · 2026
Working research paper