LiveOption: Evaluating LLM Agents in Structured Option Trading with Nonlinear Payoffs

LiveOption is a novel evaluation framework designed specifically for LLM-based agents in structured option trading, addressing the gap where existing benchmarks focus primarily on equity trading. The framework evaluates agents across three distinct scenarios: portfolio overlays, event-driven earnings trading, and 0DTE intraday trading, while enforcing realistic constraints such as margin requirements and nonlinear payoff dynamics.

Experiments reveal that current LLM agents struggle significantly in this domain, often failing to achieve competitive returns. The study highlights that option trading requires complex reasoning about volatility, time decay, and multi-leg strategy construction, rather than simple directional predictions. LiveOption provides a hierarchical metric suite to assess not just outcomes, but also action validity, risk characteristics, and decision quality, offering a more rigorous testbed for financial AI systems.

LiveOption is an evaluation framework for LLM agents in option-trading settings, motivated by the observation that many LLM multi-agent-system (MAS) trading benchmarks remain implicitly equity-centric. Such benchmarks typically measure directional prediction, single-asset buy/sell decisions, or simple portfolio returns, whereas option trading requires reasoning about nonlinear payoff functions, multi-leg positions, expiries, moneyness, volatility, and interacting risk factors. The material therefore reframes the evaluation problem: agents are not rewarded merely for forecasting prices, but for constructing or managing option structures whose realized payoff profile, risk exposure, and constraint satisfaction align with a stated objective.

The key contribution is a structured, payoff-aware testbed for LLM agents in which option strategies—such as multi-leg spreads or volatility positions—must be evaluated in terms of their payoff shape and risk behavior rather than by superficial signal accuracy. By exposing the nonlinearity and combinatorial interactions of option positions, the work highlights a gap in current LLM MAS evaluations: models that appear competent on equity-style tasks may fail to account for convexity, hedging relationships, capital or margin constraints, and the way multiple legs can offset or amplify tail risk. The implied insight is that option-trading agents require more than generic trading heuristics or natural-language reasoning; they benefit from explicit financial representations, structured constraint handling, and possibly tool-augmented pricing or risk computation.

This matters because options are a core instrument class in modern markets, central to hedging, risk management, and structured portfolio construction. If LLM agents are to be assessed as practical financial decision-makers, benchmarks must move beyond linear, single-asset proxies and test whether models can handle the nonlinear, multi-factor, and risk-managed nature of derivatives. LiveOption thus provides a more financially grounded standard for LLM MAS research, revealing failure modes hidden by equity-centric evaluations and pointing toward agent designs that better integrate option mathematics, market structure, and explicit risk controls.