SOTA: Stock Options Trading Agents Guided by Option-Implied Return Distributions

AI summary

Overview

Research area: Artificial intelligence for financial decision-making — specifically language-model-based agentic trading systems applied to listed equity options.

Technical level: Advanced. The paper assumes familiarity with reinforcement learning (policy optimization, reward design, Markov decision processes), option pricing theory (Black–Scholes–Merton Greeks, implied volatility surfaces), and large language model post-training (supervised fine-tuning, GRPO).

Scope in one sentence: The paper proposes SOTA, a language-model agent trained to select among nine structured option-strategy families rather than individual contracts, evaluated against rule-based and machine-learning baselines over a six-month out-of-sample period on options on nine large-cap U.S. equities and SPY.

What This Paper Is About

Option trading is a hard decision problem for an AI agent because a single underlying stock can have thousands of listed contracts that differ in strike, maturity, and call/put type, and the available set changes constantly as contracts expire and new ones are listed. Existing approaches either predict returns for individual contracts or restrict the policy to a single fixed payoff structure (such as a straddle or a volatility trade), which prevents the agent from switching strategies when market conditions change. SOTA's goal is to let a language model reason at the level of economically meaningful strategy families while deterministic software handles the messy contract-level implementation.

Key Contributions

1. Reformulating option trading as structured strategy selection. Rather than choosing among contracts, the agent selects one of nine strategy families and its parameters. Deterministic resolvers then map that decision into exact contracts, position sizes, and hedge trades. This collapses a contract-level decision problem into a compact space spanning directional, volatility, skewness, and curvature trades.

2. A point-in-time trading environment built on option-implied return distributions. The market state is constructed from variables drawn from the option-pricing and option-return literature — 10 features per underlying including underlying return, realized volatility, implied-volatility level and change, the implied-minus-realized volatility wedge, term structure, skewness, curvature, option flow, and changes in open interest.

3. A two-stage post-training pipeline for Qwen3.8-27B. The base model is first supervised-fine-tuned on anonymized frontier-model trading trajectories, then optimized with reinforcement learning (GRPO) directly on portfolio returns net of transaction costs.

4. An empirical and ablation study. SOTA earns an 18.32% total return with a Sharpe ratio of 1.60 and a maximum drawdown of 8.96% over a six-month out-of-sample period, while all rule-based and machine-learning baselines produce negative returns. Ablations isolate the roles of reinforcement learning and of news in the information set.

Main Findings

• SOTA is the only feasible policy with positive returns in the comparison. Over the test period, SOTA achieves a total return of 18.32%, an annualized Sharpe ratio of 1.60, an annualized Calmar ratio of 4.59, an annualized Sortino ratio of 4.77, annualized volatility of 21.52%, a maximum drawdown of 8.96%, a win rate of 29.17%, and a profit/loss ratio of 4.68.

• All rule-based and machine-learning baselines lose money in the same environment. The equal-weighted rule returns −5.22% (Sharpe −13.46, MDD 5.22%, win rate 31.11%, PLR 1.08); the GARCH rule returns −49.24% (Sharpe −6.27, MDD 49.42%); the threshold rule returns −28.46% (Sharpe −1.95, MDD 33.16%, win rate 41.31%, the best win rate among tradable policies); the GBDT policy returns −49.78% (Sharpe −8.55, MDD 49.78%); and the logistic policy returns −55.45% (Sharpe −11.63, MDD 55.47%). All policies share the same test dates, execution model, transaction-cost assumptions, portfolio constraints, and deterministic resolvers.

• A hindsight oracle bounds what the strategy space can deliver. The non-tradable oracle, which picks the best-performing strategy from the same candidate set using realized future outcomes under identical costs and constraints, returns 608.31% with a Sharpe ratio of 6.28, a Calmar ratio of 740.22, a Sortino ratio of 35.87, annualized volatility of 63.91%, and a maximum drawdown of 7.32%. Win rate and profit/loss ratio are not reported for the oracle. This reference indicates substantial profitability exists inside the structured strategy space beyond what SOTA captures.

• News plays an asymmetric role. Including anonymized news improves the quality of frontier-model trajectories used for supervised fine-tuning. However, holding the supervised checkpoint fixed and retaining contemporaneous news during reinforcement learning lowers total return from 18.32% to −2.72% and raises maximum drawdown from 8.96% to 31.69% (annualized volatility 36.00%, Sharpe −0.16, Calmar −0.17, Sortino −0.29, win rate 38.89%, PLR 1.36).

• Reinforcement learning improves on supervised fine-tuning alone under the same information set. The SFT-only policy (no RL, with news) returns −10.19% with a Sharpe ratio of −1.25, a Calmar ratio of −1.27, a Sortino ratio of −1.91, annualized volatility of 17.59%, a maximum drawdown of 15.56%, a win rate of 37.50%, and a PLR of 0.97. Adding RL raises total return to −2.72% and the Sharpe ratio to −0.16, though volatility and maximum drawdown increase. The full method's performance additionally depends on removing news during policy optimization.

• The frontier teacher's retained trajectories were positive but modest in aggregate. From 1,000 generated trajectories, those with an annualized Sharpe ratio above 0.75 were kept, yielding 144 trajectories used for supervised fine-tuning. Across this retained set, the total return was +7.99%. By family: debit vertical +2.86%, credit vertical +0.73%, outright +3.16%, defined risk reversal +0.22%, long strangle +0.68%, iron butterfly +1.54%, iron condor +2.21%, long straddle −0.57%, butterfly −0.01%, and the delta hedge component −2.83%.

Methodology in Plain English

The core trick: shrink the action space. Instead of asking the model to pick contracts out of a chain of thousands, SOTA asks it to pick a shape of trade. Each action is specified as a tuple — underlying, strategy family, long/short orientation, tenor bucket, and coordinates on a delta grid. The nine strategy families are outright, debit vertical, credit vertical, defined risk reversal, long straddle, long strangle, iron butterfly, iron condor, and butterfly, which together cover directional, volatility, skewness, and curvature exposures. The agent can open, close, roll, or hold. To avoid feeding the whole option chain into the model's context, the agent proposes up to six candidate strategies per underlying, and only those packages are priced before the final choice.

Deterministic resolvers do the bookkeeping. Three resolvers handle what the language model does not: a contract resolver picks the listed expiry closest to the tenor bucket anchor (5, 14, 45, or 120 days) and the strike whose delta exposure is closest to the requested delta; a size resolver trades in integer units for equal notional and equal risk, where risk is estimated with a second-order Taylor expansion of package value under a one-standard-deviation adverse underlying move, a 10% implied-volatility move, and one day of carry; and a hedge resolver neutralizes delta on non-directional strategies using a no-hedge band derived by Whalley and Wilmott (1997) to economize on transaction costs.

The environment is point-in-time. State combines market features, textual summaries of news and filings, open option positions, and portfolio-level information. Portfolio value is marked from options, the underlying, and cash, with trades executed at the midpoint of the best bid and ask and fees applied. Execution assumes $0.65 per contract per leg in option fees, $5 per assignment event, $0.005 per share in share commissions, a 0.5% annual stock borrow cost, a stop-loss at 60% of entry premium, and a ruin floor at 20% of initial net asset value.

Training happens in two phases. In Phase I, a frontier teacher generates trading trajectories over the supervised period starting from $1,000,000 in net asset value, with stock identity, absolute price levels, and calendar time masked to reduce temporal and identity leakage while retaining market states and contemporaneously available news. The base policy is fine-tuned on these trajectories. In Phase II, the policy is initialized from the supervised checkpoint and optimized with GRPO using a reward equal to the change in log portfolio value, directly tying optimization to portfolio performance net of transaction costs.

Data and splits. The evaluation covers ten actively traded underlyings: SPY plus nine large-capitalization U.S. equities (AAPL, AMZN, GOOGL, META, MSFT, MU, NVDA, PLTR, TSLA). Option data come from OPRA trades and quotes and OptionMetrics; underlying stock prices from CRSP daily open and close for the equities and SpiderRock marks for SPY; news from Massive.com's news API; and filings from SEC EDGAR. The chronological splits are 2024-09-03 to 2024-11-29 for frontier-model generation and supervised fine-tuning, 2024-12-02 to 2025-02-28 for reinforcement learning, and 2025-03-03 to 2025-08-29 for out-of-sample test. Performance is scored on eight metrics: total return, annualized Sharpe ratio, annualized Calmar ratio, annualized Sortino ratio, annualized volatility, maximum drawdown, win rate, and profit/loss ratio.

Why This Matters

Impact on research. The paper makes a design argument that recurs across agentic AI: the size and structure of the action space matter as much as the capabilities of the policy model. By showing that a 27B-parameter language model can beat supervised machine-learning policies that see only numerical features, it suggests that contextual reasoning adds value in options markets — but only when the interface between reasoning and execution is carefully designed. The news ablation is arguably the most provocative result: information that demonstrably helps construct expert supervision can actively hurt when left in the loop during policy optimization, which challenges the common assumption that more context is monotonically better.

Real-world applications:

• Systematic options desks and volatility funds looking for a way to switch dynamically between directional, volatility, skew, and curvature exposures without hand-coding each regime rule.
• Brokerage and wealth-management platforms that want to offer retail or advisor clients strategy recommendations at the payoff-structure level rather than the contract level, where selection is bewildering.
• Risk management and hedging, since the resolver architecture already contains a delta-hedging layer with a transaction-cost-aware no-hedge band that could be extended to a broader hedging mandate.
• Trading-agent benchmarking, as the shared environment, shared execution assumptions, and shared baselines provide a controlled testbed in which strategy-selection rules can be compared without portfolio-implementation differences confounding the result.

Industry relevance. The comparison against GBDT and logistic policies is direct: on this test period, purely supervised numerical pipelines lost between roughly half and roughly 55% of capital while the post-trained language policy gained 18.32%. For firms already experimenting with LLM agents in finance, the paper supplies a concrete architecture — structured action interface plus deterministic resolvers plus RL on net-of-cost returns — rather than a general exhortation to "use an LLM." The transaction-cost realism (per-leg fees, per-share commissions, assignment fees, borrow costs) and the explicit stop-loss and ruin-floor rules also make the setup closer to a deployable simulation than to a frictionless backtest.

Future Directions

1. Extending the learned decision space to contract selection, position sizing, and dynamic hedging. The authors explicitly name this as future work. At present these are handled by deterministic resolvers rather than learned by the policy, which caps how much of the oracle's 608.31% return could in principle be captured.

2. Understanding why news hurts during reinforcement learning. The paper documents the asymmetry — news helps the frontier teacher, hurts during policy optimization — but the mechanism is not established. Candidate explanations worth probing include reward hacking on spurious textual correlates, distribution shift between teacher and student information sets, and overfitting to news patterns that do not persist out of sample.

3. Alternative representations of structured market and textual information. The paper notes this as an open direction. The current state is 10 numerical features per underlying plus textual summaries; whether richer surface representations or differently filtered text would change the news result is unresolved.

4. Narrowing the gap to the hindsight oracle. SOTA's 18.32% return versus the oracle's 608.31% return over the same period, using the same candidate set, costs, sizing rules, limits, and hedging rule, indicates that the strategy space itself is far richer than the learned policy currently exploits. Better selection within the existing space is a substantial, well-scoped open problem.

Target Audience

This paper is most valuable to quantitative researchers and practitioners working at the intersection of large language models and financial markets — particularly those building or evaluating trading agents, and those who need to justify an architecture choice between contract-level and strategy-level decision interfaces. It will also interest reinforcement learning researchers studying post-training on domain-specific environments with real economic rewards, since the two-phase SFT-then-GRPO pipeline and the news ablation are informative beyond finance. Options traders and volatility strategists will find the resolver design (delta-grid strikes, tenor buckets, Taylor-expansion risk sizing, Whalley–Wilmott hedging bands) useful as an engineering blueprint. It is likely too specialized for general AI readers without background in either options or RL, and readers seeking deployable alpha should note that the evaluation covers a single six-month out-of-sample window on ten underlyings.

Authors’ abstract

As option markets grow and AI advances, agentic systems for option trading are gaining increasing attention. Language-model-based agents can reason over contextual information such as news, but option trading presents a particularly challenging decision problem: a single stock can have thousands of contracts, and the agent must decide both which contracts to trade and how to combine them. Existing approaches often sidestep this complexity by restricting the policy to a fixed strategy structure, such as a straddle, limiting their ability to switch strategies as market conditions change. We present SOTA (Stock Options Trading Agents), an agentic trading framework for structured option-strategy selection. SOTA abstracts the large option universe into strategy-level decisions while deterministic resolvers handle portfolio implementation. We develop SOTA by post-training Qwen3.8-27B with supervised fine-tuning followed by reinforcement learning. SOTA is evaluated on options on nine large-cap U.S. equities and SPY against rule-based and machine-learning strategy selectors in the same trading environment. Over a six-month out-of-sample period, SOTA earns an 18.3% total return with a Sharpe ratio of 1.60 and a maximum drawdown of 8.96%. We also document an asymmetric role of news: news improves frontier-teacher trajectories, but retaining news during reinforcement learning reduces out-of-sample return from 18.3% to -2.7%.