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

Yizhen Xie; Mengyang Liu

cs.AIOct 7, 2026

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

Authors: Yizhen Xie, Mengyang Liu

Organizations: Carnegie Mellon University · Amazon

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%.

Figures & tables

Figure 1: From contract-level to strategy-level decisions. Left: at the contract level, an agent must choose among thousands of listed contracts that differ in strike, expiry, and call/put type. Right: SOTA instead has the agent select one of nine option-strategy families and its parameters (here, a long straddle), and deterministic tools resolve that choice into exact contracts, a position size, and hedge trades.

Phase	Dates	Role
Supervised fine-tuning	2024-09-03 to 2024-11-29	Frontier-model generation and supervised fine-tuning
Reinforcement learning	2024-12-02 to 2025-02-28	Reinforcement learning to optimize the policy
Test	2025-03-03 to 2025-08-29	Out-of-sample evaluation of the policy and baselines

Table 1: Chronological data splits.

Rule-based policies
	Profit	Risk-adjusted profit			Risk		Trade
Policy	TR(%) $\uparrow$	ASR $\uparrow$	ACR $\uparrow$	ASoR $\uparrow$	AVOL(%) $\downarrow$	MDD(%) $\downarrow$	WR(%) $\uparrow$	PLR $\uparrow$
Equal-weighted rule	$-5.22$	$-13.46$	$-1.98$	$-11.54$	$\mathbf{0.81}$	$\mathbf{5.22}$	$31.11$	$1.08$
GARCH rule	$-49.24$	$-6.27$	$-1.52$	$-6.25$	$22.17$	$49.42$	$29.85$	$0.75$
Threshold rule	$-28.46$	$-1.95$	$-1.50$	$-3.50$	$35.20$	$33.16$	$\mathbf{41.31}$	$1.89$
Machine-learning policy

Table 2: Baseline comparison for option-strategy selection policies

Full method
	Factor		Profit	Risk-adjusted profit			Risk		Trade
Variant	RL	News	TR(%) $\uparrow$	ASR $\uparrow$	ACR $\uparrow$	ASoR $\uparrow$	AVOL(%) $\downarrow$	MDD(%) $\downarrow$	WR(%) $\uparrow$	PLR $\uparrow$
SOTA	✓	$\times$	$\mathbf{18.32}$	$\mathbf{1.60}$	$\mathbf{4.59}$	$\mathbf{4.77}$	$21.52$	$\mathbf{8.96}$	$29.17$	$\mathbf{4.68}$
Ablations
w/ news during RL	✓	✓	$-2.72$	$-0.16$	$-0.17$	$-0.29$	$36.00$	$31.69$	$\mathbf{38.89}$	$1.36$
w/o RL (SFT only)	$\times$	✓	$-10.19$	$-1.25$	$-1.27$	$-1.91$	$\mathbf{17.59}$	$15.56$	$37.50$	$0.97$

Table 3: Ablation studies over the test period. Bold marks the best value in each column.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Table 4 — Table 4: Expiration profit and loss of the nine option-strategy families, net of the initial cost of the package a

Assumption	Value
Traded price	Midpoint of the best bid and best ask
Option fee	$0.65/contract/leg
Assignment fee	$5/event
Share commission	$0.005/share
Stock borrow	0.5% annually
Stop-loss	60% of entry premium

Appendix

Table 5: Trading and execution assumptions.

Directional
			Profit	Risk-adjusted			Risk
Code	Family	Trades	TR(%)	ASR	ACR	ASoR	AVOL(%)	MDD(%)	WR(%)
ol	Outright	4.29	$+3.16$	1.43	5.08	2.93	6.91	2.33	46.3
dv	Debit vertical	2.84	$+2.86$	2.10	8.24	4.60	5.51	1.68	75.3
cv	Credit vertical	1.46	$+0.73$	1.97	14.88	7.28	2.54	0.70	82.4
dg	Defined risk reversal	0.28	$+0.22$	1.31	4.48	1.99	4.87	1.51	60.9

Appendix

Table 6: Frontier teacher performance by strategy families.

a_{t}^{T}\sim\pi_{T}(\cdot\mid s_{t},n_{t})

Appendix

Algorithm 1 SOTA Training

Parameter	Symbol	Value	Selection method
ARCH coefficient	$\alpha$	0.060	QML on estimation sample
GARCH coefficient	$\beta$	0.762	QML on estimation sample
Persistence	$\alpha+\beta$	0.822	implied (half-life 3.5 d)
Winsorisation	—	$4\bar{\sigma}_{i}$	fixed a priori
Forecast horizon	$H$	10 trading days	matches the 14-day tenor anchor
Estimation sample	—	3,710 obs, 2023-09-01 to 2025-02-28	strictly pre-test

Appendix

Table 7: GARCH rule.

Family	Code	Feature	Entry	Edge	Friction
Participation gate, applied to every family
all	—	$\mathrm{ts}$	$<0$	—	—
Volatility: $\mathrm{iv}/\mathrm{rv}$ , fair value $1.0$
Long straddle	ls	$\mathrm{iv}/\mathrm{rv}$	$\leq 0.90$	10%	3.2%
Iron butterfly	ib	$\mathrm{iv}/\mathrm{rv}$	$\geq 1.20$	20%	9.1%
Iron condor	ic	$\mathrm{iv}/\mathrm{rv}$	$\geq 1.25$	25%	10.6%

Appendix

Table 8: Threshold rule.

Model and supervised fine-tuning
Parameter	Qwen3.8-27B
Hugging Face revision	1d4bf0f2
Precision	bf16
SFT episodes	$414$
Maximum sequence length	$33{,}280$
Train batch size	$16$

Appendix

Table 9: Training hyperparameters for Qwen3.8-27B.

Parameter	Symbol	Value
Universe and action space
Underlyings	$N$	10 (nine single names, one index slot)
Initial net asset value	$V_{0}$	$1,000,000
Cash accrual	$r_{f}$	0
Registered families	$\|\mathcal{F}\|$	9
Tenor buckets (days to expiry)	$\mathcal{K}$	0–7, 8–30, 31–90, 91–180

Appendix

Table 10: Action space, sizing and hedging parameters.

Family	Default coordinates	Admitted range
Outright	long delta 0.55	$[0.20,0.80]$
Debit vertical	long delta 0.55, width 0.10	long $[0.20,0.80]$ , width $[0.05,0.30]$
Credit vertical	short delta 0.35, width 0.10	short $[0.10,0.55]$ , width $[0.05,0.30]$
Butterfly	centre 0.45, widths 0.10/0.10	centre $[0.25,0.70]$ , widths $[0.05,0.30]$
Defined-risk reversal	directional 0.35, tail wing 0.20	directional $[0.15,0.60]$ , wing $[0.05,0.35]$
Long straddle	call 0.50, put 0.50	$[0.10,0.55]$

Appendix

Table 11: Family default coordinates and admitted ranges.

Explore similar work

LiveOption: Evaluating LLM Agents in Structured Option Trading with Nonlinear Payoffs
Sep 27, 2026Haochen Luo, Yifan Li, Binh Minh An +4Algorithmic Trading Large Language Model Agents
City University of Hong Kong · The Chinese University of Hong Kong (Shenzhen) · Shanghai University of Finance and Economics
EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents
Sep 15, 2026Sehee Kim, Yumin Choi, Minki Kang +1On-Policy Self-Evolution Alphaevolve
KAIST
Strat-LLM: Stratified Strategy Alignment for LLM-based Stock Trading with Real-time Multi-Source Signals
May 7, 2026Wenliang Huang, Zengyi YuLarge Language Model Alignment Strategic Reasoning
School of Economics Zhejiang University of Technology Hangzhou, China · Faculty of Education East China Normal University Shanghai, China

Sep 27, 2026, cs.AIJ/K move · Enter open · S save

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

Haochen Luo, Yifan Li, Binh Minh An, Xiaolong Luo+3

City University of Hong Kong · The Chinese University of Hong Kong (Shenzhen) · Shanghai University of Finance and Economics

Algorithmic Trading Large Language Model Agents

Open paper AskPDF arXiv

Large language models (LLMs) and multi-agent systems (MAS) have shown promise in financial decision-making, yet existing evaluations focus on equity trading and primarily assess directional prediction, overlooking the structural complexity of derivative markets. Option trading introduces fundamentally different challenges, including nonlinear payoffs and multi-leg strategy construction, requiring structured decisions rather than simple directional bets. We introduce LiveOption, an evaluation framework for LLM-based agents in option trading. LiveOption formulates the problem as structured sequential decision-making under realistic execution and capital constraints, and provides a reproducible environment with standardized interaction protocols. The framework includes three task suites covering portfolio overlays, event-driven earnings trading, and 0DTE intraday trading. We further propose a hierarchical metric suite that evaluates action validity, decision quality, risk characteristics, and outcome-level performance. Experiments show that current agents often fail to achieve competitive returns in most scenarios. LiveOption offers a principled testbed for evaluating structured decision-making beyond outcome-based metrics.

Abstract

Explore similar work

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

EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents

Strat-LLM: Stratified Strategy Alignment for LLM-based Stock Trading with Real-time Multi-Source Signals