Paper: arXiv 2512.01123
Authors: Xiaoting Kuang, Boken Lin
Abstract
Large Language Models (LLMs) excel at understanding context and qualitative nuances but struggle with the rigorous and transparent reasoning required in high-stakes quantitative domains such as financial trading. We propose a model-first hybrid architecture for the options “wheel” strategy that combines the strengths of LLMs with the robustness of a Bayesian Network. Rather than using the LLM as a black-box decision-maker, we employ it as an intelligent model builder. For each trade decision, the LLM constructs a context-specific Bayesian network by interpreting current market conditions, including prices, volatility, trends, and news, and hypothesizing relationships among key variables. The LLM also selects relevant historical data from an 18.75-year, 8,919-trade dataset to populate the network’s conditional probability tables. This selection focuses on scenarios analogous to the present context. The instantiated Bayesian network then performs transparent probabilistic inference, producing explicit probability distributions and risk metrics to support decision-making. A feedback loop enables the LLM to analyze trade outcomes and iteratively refine subsequent network structures and data selection, learning from both successes and failures. Empirically, our hybrid system demonstrates effective performance on the wheel strategy. Over nearly 19 years of out-of-sample testing, it achieves a 15.3% annualized return with significantly superior risk-adjusted performance (Sharpe ratio 1.08 versus 0.62 for market benchmarks) and dramatically lower drawdown (-8.2% versus -60%) while maintaining a 0% assignment rate through strategic option rolling. Crucially, each trade decision is fully explainable, involving on average 27 recorded decision factors (e.g., volatility level, option premium, risk indicators, market context).
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 9.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced probabilistic graphical models (Bayesian Networks) and quantitative risk metrics (Sharpe ratios, drawdowns) but grounds them in a comprehensive 19-year backtest with specific performance data, showing strong empirical execution.
Research Flowchart
flowchart TD A["Research Goal: Hybrid AI for Options Wheel Trading<br>LLM understanding + Quantitative rigor"] --> B["Input: Market & Historical Data<br>Market conditions, prices, volatility, trends, news<br>18.75-year dataset: 8,919 trades"] B --> C["LLM-Generated Bayesian Network<br>Interprets context & builds probabilistic model<br>Selects analogous historical data"] C --> D["Transparent Probabilistic Inference<br>Generates explicit probability distributions<br>Calculates risk metrics (27 decision factors)"] D --> E["Trade Decision & Execution<br>Strategic option rolling to manage risk"] E --> F["Feedback Loop & Performance<br>Outcomes analyze network accuracy & refine structure"] F --> G["Key Findings: Superior Performance<br>15.3% Annualized Return (19-year OOS)<br>Sharpe Ratio 1.08 vs 0.62 market<br>Max Drawdown -8.2% vs -60%<br>0% Assignment Rate"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a model-first hybrid architecture that uses LLMs to generate context-specific Bayesian networks for each options wheel trade decision.
Source sentence
We propose a model-first hybrid architecture for the options "wheel" strategy that combines the strengths of LLMs with the robustness of a Bayesian Network.
Assumes 4
- The LLM-generated Bayesian network structure is appropriate for the given market context. model
Source sentence
For each trade decision, the LLM constructs a context-specific Bayesian network by interpreting current market conditions—including prices, volatility, trends, and news—and hypothesizing relationships among key variables.
- Historical trade data from Jan 2007–Dec 2015 is sufficient to train the network without leakage. data
Source sentence
We train the Bayesian network structure generation and initial probability estimation solely on data from January 2007 to December 2015 (9 years).
- Positions are limited to a maximum of 5% of average daily volume to control market impact. execution
Source sentence
Our position sizing methodology accounts for market impact by limiting individual positions to a maximum of 5% of the average daily volume.
- Transaction costs are modeled as fixed commissions and average slippage rates. other
Source sentence
Each options trade incurs a commission of $0.65 per contract plus a $0.10 per contract exchange fee, with a minimum commission of $1.00 per trade
Data
2007 – 2025-09 · equity options · options
- Wheel strategy trades dataset options public
Source sentence
Our validation employs a comprehensive dataset of wheel strategy trades from January 2007 to September 2025, covering 18.75 years of real market data. This dataset includes 8,919 individual trades executed with an initial capital of $100,000.
- Historical trade database options public
Source sentence
Instead of using LLM-generated probabilities, the system queries our database of 8,919 documented historical trades, each characterized by 27 decision factors, to populate conditional probability tables.
Costs
Modeled
Commission $0.65 per contract + $0.10 exchange fee (minimum $1 per trade); total commissions ≈ $12,518. Slippage 0.15% of premium for puts, 0.12% for calls. Position sizing limited to 5% of average daily volume.
Source sentence
Our position sizing methodology accounts for market impact by limiting individual positions to a maximum of 5% of the average daily volume.
Finds 3
Sharpe 1.08 (benchmark 0.62) · annual return 15.3% · max drawdown 8.2%
- The hybrid system achieved a 15.3% annualized return with a Sharpe ratio of 1.08 and a maximum drawdown of -8.2% over 18.75 years of out‑of‑sample testing. performance
Source sentence
Over nearly 19 years of out-of-sample testing, it achieves a 15.3% annualized return with significantly superior risk-adjusted performance (Sharpe ratio 1.08 versus 0.62 for market benchmarks) and dramatically lower drawdown (-8.2% versus -60%) while maintaining a 0% assignment rate.
- After accounting for transaction costs, net annualized return is 15.0%, a reduction of about 0.3 percentage points. performance
Source sentence
These transaction costs collectively reduce annualized returns by approximately 0.3 percentage points, resulting in net performance of 15.0% after costs, compared to 15.3% gross returns.
- The hybrid approach outperforms baselines: pure LLM 8.7% return (Sharpe 0.45), static Bayesian 11.2% (Sharpe 0.67), rules‑based 9.8% (Sharpe 0.52). performance
Source sentence
Pure LLM approaches yield 8.7% returns with a 0.45 Sharpe ratio. Static Bayesian networks achieve 11.2% returns and a 0.67 Sharpe ratio. Rules-based systems produce 9.8% returns with a 0.52 Sharpe ratio. In contrast, our hybrid approach attains 15.3% returns and a 1.08 Sharpe ratio.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 23 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.