PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Haochen Luo, Yifan Li, Binh Minh An, Xiaolong Luo, Zhengzhao Lai, Yuan Zhang, Chen Liu

arXiv:2609.33470v2cs.AIq-fin.CP

TL;DR

Existing trading-agent benchmarks largely focus on directional decisions in equities, leaving option trading’s nonlinear, constraint-rich decisions without a comprehensive evaluation framework. LiveOption provides a multi-scenario benchmark and options-specific evaluation, and its experiments find negative typical episode returns across tasks. The authors argue for domain-specific benchmarks and reporting distributions and interventions rather than means alone.

  • Problem

    Existing agent evaluations focus mainly on equity trading and directional prediction, while options require structured decisions involving nonlinear payoffs, multi-leg strategies, and constraints.

    Source: §1 · p. 2

  • Method

    LiveOption evaluates LLM option-trading agents across overlay, earnings, and 0DTE scenarios using realistic options backtesting and outcome and behavioral metrics.

    Source: §1 · p. 2, §2.4 · p. 4, §2.5 · p. 5

  • Results

    Typical episode returns are negative in every task, only the widest earnings gap survives multiple-comparison correction, and removing portfolio state roughly trebles mean maximum drawdown.

    Source: §5 · p. 10

  • Takeaways & Limitations

    The authors argue for domain-specific benchmarks beyond directional prediction and for reporting distributions and interventions rather than means alone.

    Source: §5 · p. 10

    The evaluation covers 2025 data, SPY 0DTE sessions, five Overlay assets, and 24 Earnings single names, so results should not be extrapolated to other sectors, asset classes, or market regimes without further evaluation.

    Source: B Limitations · p. 17

Abstract

from arXiv · show

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. Project Website: https://liveoption.github.io/

1 Introduction

Existing trading-agent evaluations largely target equities or cryptocurrencies and simple directional actions, while option markets add nonlinear payoffs, time decay, volatility sensitivity, margin constraints, and multi-dimensional risk. LiveOption addresses the resulting evaluation gap with a multi-scenario benchmark and realistic execution for LLM-driven option agents.

1 Introduction

  • Options add nonlinear payoffs, time decay, volatility sensitivity, margin constraints, and strategy choices across strikes and maturities, making agent decisions structurally different from directional equity trading.Risk must also be managed across multiple Greeks rather than price direction alone.Source: §1 · p. 2
  • LiveOption evaluates LLM-driven agents across portfolio overlays, earnings-event trading, and intraday 0DTE under evolving market conditions and strict risk and margin constraints.Its backtesting engine simulates brokerage-style execution, margin enforcement, and multi-leg position management.Source: §1 · p. 2
  • The work presents what it describes as the first systematic study of LLM-based trading agents in option markets, addressing derivative-specific structural challenges.It establishes a foundation for future research in this domain.Source: §1 · p. 2
  • The benchmark formalizes option-trading scenarios and supports scalable user-defined scenario construction beyond its predefined tasks.Its scenarios cover portfolio overlay, earnings-event-driven trading, and intraday 0DTE trading.Source: §1 · pp. 2–3
  • Behavioral analysis examines agent decisions, risk-management discipline, and execution quality beyond outcome-level metrics.Experiments across state-of-the-art LLMs identify systematic option-trading failure modes and derive insights for future agent design.Source: §1 · p. 3

2 Framework of LiveOption

LiveOption combines live option-market contexts, LLM agents, an options-specific backtesting engine, and multi-level evaluation. Its design spans distinct trading tasks, constrains agent actions to executable orders, and evaluates both outcomes and decision behavior.

2.1 Overall Design

  • The overall design combines a live environment, a multi-agent system, a backtesting engine, and an evaluation suite.The live environment specifies available information, tasks, and user configurations.Source: §2.1 · p. 3, Figure 2 · p. 3

2.2 Live Environments

  • LiveOption defines three environments with different trading frequencies, investment styles, and rules: portfolio overlays, earnings bets, and 0DTE intraday trading.The environments represent practical option use cases and investment intentions.Source: §2.2 · pp. 3–4
  • The environment can be customized with different objectives, portfolio constraints, and trading frequencies for targeted evaluation settings.Examples include return maximization, risk control, income generation, capital allocation, leverage limits, and intraday, daily, or event-driven trading.Source: §2.2 · p. 4

2.3 LLM-based Trading Agents

  • The analyzer–executor design separates high-level market interpretation and trading decisions from order construction under real-world constraints.This adapts a generic agent-system paradigm to option trading’s high-dimensional data and constrained combinatorial actions.Source: §2.3 · p. 4
  • The framework reduces option-chain complexity through a compact volatility representation and a discrete structured action space, then verifies orders against trading constraints.Backend checks enforce margin, position limits, and contract validity.Source: §2.3 · p. 4

2.4 Backtesting Engines

  • The framework pairs scenario replay, which reconstructs market conditions, with a trading simulator that maintains brokerage-style account state and enforces realistic constraints.The design is tailored to options’ contract-aware mechanics and rich market context.Source: §2.4 · p. 4

2.5 Evaluation Suites

  • Evaluation reports portfolio-level returns, risk, and benchmark-relative metrics for overlays, and episode-level outcomes for earnings and 0DTE tasks.It also uses resampling intervals and paired tests for uncertainty and comparisons.Source: §2.5 · p. 5
  • Behavioral evaluation uses decision traces to analyze strategy composition, directional versus structural reasoning, position sizing, leverage, and Greek contributions to realized PnL.The analysis distinguishes option-mechanics considerations from directional views and incidental directional gains.Source: §2.5 · p. 5

3 Benchmarks, Analysis and Discussion

Across three option-trading scenarios, agents generally fail to produce reliable returns, while behavioral analysis and targeted input ablations reveal differences that outcome comparisons alone miss. The analysis covers a 2025 evaluation of eight models and motivates risk checks, richer reporting, and intervention-based testing.

3.1 Experimental Settings

  • The evaluation uses 2025 market data and eight LLMs; users can also import their own market and news data through documented schemas.The authors selected 2025 to limit pretraining information leakage for models whose knowledge cutoffs preceded that year and to include diverse market regimes.Source: §3.1 · pp. 5–6

3.2 Overall Performance

  • Overlay active returns are mostly negative, all episodic-task medians are negative, and no 0DTE pairwise model difference survives Holm correction; earnings results do not establish a reliable ranking.The tables report overlay performance against matched buy-and-hold and episode-level performance for Earnings and 0DTE.Source: Table 1 · p. 6, §3.2 · pp. 6–7, Table 2 · p. 7

3.3.1 Overlay Behavior: Instrument Choice, Tenor and Exposure

  • Overlays mostly trail buy-and-hold but beat fixed 0.30-delta covered-call rules by 17.5 active-return points; tenor varies by model, with sustained exposure drift confined to one case.The reported exposure drift is for Llama-3.3-70B-Instruct under the hedge mandate.Source: §3.3.1 · pp. 7–8

3.3.2 Action Concentration Within Tasks

  • Trading activity varies two- to fivefold between models on the same task, while structure concentration differs between 0DTE and Earnings rather than following one fixed template.Income fills range from 164 to 852; single-leg positions comprise 49.2%–96.1% of classified 0DTE structures, while vertical and two-leg positions comprise 43.3%–89.1% in Earnings.Source: §3.3.2 · p. 8, Table 4 · p. 8, Figure 4 · p. 8

3.3.3 Episode-Return Distributions and Tail Risk

  • Both episodic return distributions are right-skewed, with every model losing on most 0DTE sessions and having a negative median in both tasks.Across the tasks, the panel records 38 upside tails against five downside tails; GPT-OSS-120B’s ten upside 0DTE tails lift its mean above zero despite a −3.20% median.Source: §3.3.3 · p. 8

3.3.4 Where Risk Is Carried: Greek Attribution

  • Earnings attribution shows large negative theta and small positive vega; in 0DTE, delta spans zero and the residual is negative for every model.The authors describe attribution as carried exposure, not evidence that exposure was mispriced.Source: Table 5 · p. 9, §3.3.4 · p. 9

3.4 Targeted Ablations

  • Removing Greeks and multi-timeframe inputs cuts DeepSeek-V4-Flash fills per session from 29.5 to 4.4 and raises its no-trade rate to 95.8%; removing portfolio state raises drawdown for every model.No return effect is significant at the 5% level after Holm correction; removing portfolio state raises mean maximum drawdown from 4.4% to 12.4%.Source: Table 6 · p. 9, §3.4 · p. 9

3.5 Implications for Robust and Reliable Agent Design in Option Trading

  • The authors recommend backend risk checks, reporting participation and drawdown beside returns, and testing information dependence with matched input-removal interventions.The recommended checks include maximum loss per trade, portfolio-level Value-at-Risk, margin feasibility, and gamma exposure.Source: §3.5 · p. 10

4 Related Works

Prior LLM trading work largely targets stocks and cryptocurrencies, while option-specific LLM research remains limited. Existing dynamic trading benchmarks also focus mainly on equities, leaving option trading under-evaluated.

LLMs in Financial Trading

  • TradingGPT [53], FinMem, CryptoTrade, Flag-Trader [55], and FinCon [16] address trading, but primarily in stocks or cryptocurrencies with simpler action spaces and linear payoffs.OQL [41] explores LLM-generated option strategies, but the application of LLM agents to options remains largely underexplored given nonlinear payoffs and complex strategy spaces.Source: §4 · p. 10

Benchmarking LLMs in financial Domain

  • Dynamic trading benchmarks such as InvestorBench [21] and StockBench [58] evaluate complex decisions, but mainly in equities; recent live benchmarks [17] also overlook options.Earlier benchmarks focus on financial document understanding and QA, while trading benchmarks have also faced prior-knowledge leakage and unrealistic assumptions such as ignoring transaction costs [23].Source: §4 · p. 10

5 Conclusion

LiveOption evaluates LLM agents across portfolio overlays, earnings trading, and intraday 0DTE options, documenting both performance and behavioral patterns. Results show negative typical episode returns across tasks and identify limitations in current evaluation coverage.

5 Conclusion

  • Typical episode returns are negative in every task, only the widest earnings gap survives multiple-comparison correction, and removing portfolio state roughly trebles mean maximum drawdown.The conclusion argues for reporting distributions and interventions rather than means alone.Source: §5 · p. 10
  • LiveOption benchmarks LLM agents across intraday 0DTE trading, earnings-driven strategies, and portfolio overlays.It is framed as a testbed for structured, nonlinear, constraint-rich option decisions.Source: §5 · p. 10

A Broader Impacts

  • Observed behavior includes widely varying tenor policies, task-specific action structures, 7.0%–66.1% 0DTE risk-rejection rates, and roughly doubled-to-quintupled mean maximum drawdown when portfolio state is removed.The paper recommends execution-layer risk verification, reporting participation and tail risk alongside return, and testing information dependence through intervention.Source: A Broader Impacts · p. 17

B Limitations

The paper outlines option mechanics and a Greek-based PnL decomposition, while noting limits on the evaluation universe, rollout sampling, execution-cost coverage, and model coverage. Its decomposition is intended to distinguish sources of realized profit and loss, with a residual that can include higher-order effects and model mismatch.

B Limitations

  • Each model–case configuration uses one selected rollout, so uncertainty estimates reflect market-episode variation rather than repeated-generation variability, and valid completed runs exclude end-to-end API reliability.The benchmark is limited to 2025 data, SPY 0DTE sessions, five Overlay equities, and 24 Earnings single names; results should not be extrapolated to other sectors, asset classes, or regimes without further evaluation.Source: B Limitations · p. 17
  • Execution costs receive detailed modeling only for 0DTE because comparable quote coverage is unavailable for Overlay and Earnings; fixed-decision repricing omits adaptation to latency, queue priority, and market impact.Computational constraints also leave smaller open-weight models and alternative agent architectures unexplored.Source: B Limitations · p. 17

C.1 Introduction of Option

  • An option gives its holder the right, but not the obligation, to trade an underlying at a pre-specified strike within a pre-specified time window.Calls grant the right to buy; puts grant the right to sell.Source: §C.1 · p. 17

C.2 Option Pricing Model

  • Option value comprises intrinsic value and time value, with the latter reflecting uncertainty and remaining time and decaying rapidly for short-maturity contracts.The BSM framework supplies a tractable reference for option pricing and motivates Greeks, while implied volatility normalizes comparisons across strikes and expiries.Source: §C.1 · p. 17, §C.2 · p. 18

C.3 Greeks of Options

  • Greeks describe option-value sensitivities to risk factors, while expiration payoffs and strategy combinations shape the trade-offs among direction, risk, and cost.Short options have limited upside and potentially large losses; spreads can bound risk, while protective puts hedge downside and straddles or strangles express volatility views.Source: §C.3 · p. 18, Table 8 · p. 19, §C.5 · p. 20

C.6 Option PnL Decomposition via Greeks

  • Greek-based PnL decomposition separates realized PnL into risk-factor contributions, but its residual absorbs higher-order terms, discrete-rebalancing error, and mismatch between Black–Scholes Greeks and market dynamics.A large residual relative to explained components can signal coarse evaluation steps or regimes where a first-order expansion is insufficient; the decomposition is aggregated across legs for multi-leg strategies.Source: §C.6 · p. 20

D.1 Dataset Preparation Pipeline

The paper describes a data-and-execution framework and three option-trading scenarios spanning portfolio overlays, earnings events, and same-day 0DTE trades. Each scenario imposes different horizons, strategy choices, and decision demands.

D.1 Dataset Preparation Pipeline

  • QFinZero supplies standardized access to provider or user-imported market and news data, which are normalized and time-aligned for decision-time queries; restricted source data are not included in the public release.Substituting data sources allows use of researchers’ own assets but does not guarantee identical numerical results.Source: §D.1 · p. 21

D.2 Design of Framework

  • LiveOption combines a Data Portal, Paper Money Simulator, and Backtesting Engine to support reproducible option-agent evaluation.The engine provides timestamp-bounded context and synchronizes actions with simulated execution, while the simulator tracks account state and option lifecycle events.Source: §D.2 · pp. 21–22

D.3 Web Interface

  • The web interface configures and monitors experiments and visualizes trading metrics; support for user-defined contexts, prompts, and asset universes is planned for future versions.Its views use the benchmark’s recorded artifacts for finished and active runs.Source: §D.3 · p. 22

E.1 Takeaway

  • The benchmark spans portfolio overlays, earnings-driven strategies, and 0DTE intraday trading, covering different horizons and volatility regimes.Overlay holds a fixed equity base while layering options; its income and hedging subtasks use a fixed cash buffer, weekly decisions, and distinct short-option versus protective-structure constraints.Source: §E.1 · p. 22, §E.2.1 · p. 24, §E.2.2 · p. 24

E.3.2 Settings

  • Earnings trading is organized into betting, reaction, and aftershock phases, requiring decisions around expected movement, volatility changes, and post-event position management.The strategy set includes directional calls and puts, vertical spreads to reduce premium and IV-crush sensitivity, and straddles or strangles for direction-agnostic volatility positions; the test uses 95 independent 2025 cases, with 90–95 valid per model.Source: §E.3.1 · p. 25, §E.3.2 · p. 25, §E.3.3 · p. 25

E.4.2 Settings

  • 0DTE options combine elevated gamma and accelerating theta decay, and the scenario divides trading into early-session positioning, intraday management, and expiration-horizon resolution.Positions must be resolved within the session; management includes reacting to intraday moves and catalysts while controlling gamma exposure near expiration.Source: §E.4.1 · p. 26, §E.4.2 · p. 26

E.4.3 Experiment

The 0DTE experiment uses 102 SPY sessions on Wednesdays and Fridays, while the overlay underlyings span distinct return, risk, and tail profiles.

E.4.3 Experiment

  • The 0DTE evaluation covers 102 independent Wednesday and Friday sessions in 2025, selected to represent heavily used expiration days and varied intraday conditions.SPY is used because it offers daily expirations, allowing same-day catalyst positions without overnight exposure.Source: §E.4.3 · pp. 26–27
  • Sessions run from 9:30 AM to 4:00 PM ET, with intraday price and option-chain observations and sub-scenarios around scheduled releases and liquidity shocks.The scenarios are designed to test responsiveness to rapidly changing conditions.Source: §E.4.3 · p. 27

E.5.1 Selected Assets for Overlay Scenario

  • Overlay underlyings range from +18.57% (TSLA) to +77.54% (AMD), versus +20.40% for QQQ, a roughly 59-percentage-point spread.The rebased curves do not collapse into one bundle, indicating distinct per-name return processes.Source: §E.5.1 · p. 27, Table 12 · p. 27
  • Annualized volatility ranges from 23.62% (QQQ) to 63.31% (TSLA), while maximum drawdown ranges from −22.88% to −48.19%.The five US large-cap technology or AI-adjacent names also differ in realized Sharpe and higher-moment characteristics.Source: §E.5.1 · p. 27, Table 12 · p. 27

E.5.2 Analysis of Earning Events

The earnings sample spans substantial volatility and reaction-pattern diversity, while the evaluation uses separate portfolio-level and episode-level metrics for different task horizons. Intraday statistics characterize SPY sessions, and uncertainty is assessed with paired resampling and tests.

E.5.2 Analysis of Earning Events

  • 18 of 24 earnings constituents (75%) have at least one event move of ≥10%, and the largest positive and negative prints are +27.09% and −30.71%.The realized cross-sectional dispersion is roughly 58 percentage points.Source: §E.5.2 · p. 27
  • The 24-name earnings universe includes 2 persistent-drift, 11 single-reversal, and 11 whipsaw names; 9 also have a tail-event-dominated quarter.Close-to-close reactions are net-positive for 11 names and net-negative for 13, spanning volatility expansion and collapse, trends, mean reversion, and asymmetric tails.Source: §E.5.2 · pp. 27–28

E.5.3 Intraday Statistics

  • Daily return is the log ratio of the session’s last close to its first open, while daily volatility uses the Garman–Klass estimator on 5-minute OHLC bars.The estimator is aggregated across the session and uses intraday range information.Source: §E.5.3 · pp. 28–29

F.1 Common Metrics

  • Overlay tasks are evaluated as long-horizon portfolios, while each 0DTE day or earnings window is an independent episode with outcomes aggregated across episodes.The framework reports portfolio-level absolute and benchmark-relative measures, and episode-level Win%, PnL%, P/L, and PF.Source: §F.1 · p. 29
  • Relative overlay metrics compare the agent portfolio with buy-and-hold because option-leg exposures can make combined-portfolio Sharpe an unclear signal of overlay quality.Cumulative Delta Return measures the overlay’s incremental cumulative PnL; Information Ratio measures active return relative to tracking error.Source: §F.1 · p. 30
  • Episode metrics distinguish profitability frequency, average return, payoff shape, and aggregate profit efficiency; PF above 1 indicates aggregate net profitability.Means and drawdowns use 10,000-resample percentile bootstrap intervals, and paired comparisons use Monte Carlo sign-flip tests with Holm correction.Source: §F.1 · pp. 30–31

F.2 Behavior Analysis

LiveOption supplements outcome metrics with decision-trace analysis and presents compact representations of market information to agents. The evaluated-model panel has a look-ahead caveat because some knowledge cutoffs cover the evaluation window.

F.2 Behavior Analysis

  • Complete thesis, intent, fill, and Greek traces support behavioral audits beyond aggregate PnL, covering strategy composition, reasoning type, and leverage control.Strategy composition is interpreted alongside episode outcomes; directional-versus-structural classification distinguishes directional views from explicit option-mechanic reasoning.Source: §F.2 · p. 31
  • Leverage-control analysis tracks capital deployed and gross notional, including leverage stacking from rolling, pyramiding, or wing-overwriting.The paper identifies this mechanism as the proximate cause of sub-−100% earnings returns documented in Section 3.Source: §F.2 · p. 32

F.3 Details of Evaluated LLMs

  • The 2025 evaluation window is not a contamination control: observation gating limits future information within a run but cannot rule out pretraining recall.Results for models whose reported cutoffs cover the window should be read with that possibility open.Source: §F.3 · p. 32

G.1 Anchor-Based Volatility Surface Representation (AVSR)

  • AVSR replaces the full irregular implied-volatility surface with nine anchors: three moneyness levels and three maturities, forming a fixed 3 × 3 summary.The representation captures volatility level, skew, and term structure while remaining interpretable and suitable for the agent.Source: §G.1 · pp. 32–33, Figure 7 · p. 33

G.2 Quantized Option Chain Representation (QOCR)

QOCR replaces a large, irregular option chain with discrete moneyness and expiry buckets, representative contracts, and a fixed-size structured action space.

G.2 Quantized Option Chain Representation (QOCR)

  • QOCR discretizes the large, irregular, and redundant option chain into a structured action space to improve tractability for LLM agents.Its moneyness buckets represent OTM puts, ATM, and OTM calls; expiry buckets cover 0–14, over 14–45, and over 45–120 days.Source: §G.2 · pp. 33–34
  • A representative contract is selected from each bucket using strike and maturity proximity, with a weighting parameter balancing the two.The figure describes outer-edge representatives for wings and the strike closest to 1.0 for ATM, with the longest expiry as a tiebreaker.Source: §G.2 · p. 34, Figure 8 · p. 35
  • An action specifies the underlying, option type, moneyness bucket, expiry bucket, and quantized position size, with multi-leg strategies extending the action structure.Contracts outside the buckets are discarded.Source: §G.2 · p. 34, Figure 8 · p. 35
  • The quantized formulation makes the action space fixed and independent of the underlying option-chain size.Discretized buckets remove unnecessary granularity and provide a consistent interface across assets.Source: §G.2 · p. 34

H.1.1 Overlay Strategies

The overlay analysis compares model portfolios with matched buy-and-hold benchmarks and fixed 0.30-delta rules, then examines uncertainty, trading styles, and stock exposure. Results vary by mandate, and some measurements have important scope limits.

H.1.1 Overlay Strategies

  • Table 17 reports full-year 2025 returns, risk-adjusted metrics, and option-flow statistics against matched buy-and-hold benchmarks for eight models.∆Ret captures the overlay’s full economic effect against the same target share count; positive ∆MDD means a shallower drawdown than the benchmark.Source: Table 17 · p. 35, §H.1.1 · p. 35
  • Mean option fills per month range from 13.7 to 71.0, while median entry maturity spans 2.5 to 28 days across six models.Only 20.2% of overlay fills are shorter than a week, and five of six models fall within their mandate’s prescribed maturity band at the median.Source: §H.1.1 · p. 37
  • Every one of the 80 overlay cases ends the year at its target share count, and 76 are below target on at most eight trading days.The four exceptions all belong to Llama-3.3-70B-Instruct; its hedge stock leg is incomplete for substantial portions of the year on four assets.Source: §H.1.1 · p. 37
  • One year of daily data on five assets does not identify a risk-adjusted edge in either direction, so the study draws no cross-model conclusion from Sharpe ratios.Only one of 80 intervals excludes zero; the active-return and drawdown columns carry the overlay findings instead.Source: §H.1.1 · p. 38
  • Neither mandate’s panel-average active return is positive; against fixed rules, agents lead by 17.6 percentage points under covered calls but are nearly tied under hedging.The hedge rule reduces drawdown roughly twice as much as the agents; beating a fixed rule does not establish an advantage over passive buy-and-hold.Source: §H.1.1 · p. 38, Table 18 · p. 38

H.1.2 0DTE Intraday

The 0DTE intraday analysis finds substantial variation across models in yearly outcomes, strategy-bucket results, and trading patterns. Greek attribution and exact repricing characterize PnL, while infrequent mark updates limit the Earnings-track attribution comparison.

H.1.2 0DTE Intraday

  • Across 102 sessions, three models finish positive at +14.4, +14.4, and +11.0 percentage points, while five finish negative and monthly results show no stable ranking.Daily win rates span 25.5% to 47.1%; no month is positive for the whole panel or negative for all of it.Source: §H.1.2 · p. 39, Table 19 · p. 39
  • Long calls are the clearest loss center at −21.4k dollars pooled, while long puts, short puts, and short calls return +5.5k, +3.6k, and −3.9k.No strategy bucket dominates for every model.Source: §H.1.2 · p. 41, Table 20 · p. 41
  • Fills per session range from 14.7 to 33.6, and most models front-load trading while inventory peaks later, indicating morning round-tripping rather than accumulation.All times are Eastern; a fixed UTC offset would misplace the 66 daylight-saving sessions by one hour.Source: §H.1.2 · p. 42, Figure 9 · p. 43
  • Table 21 decomposes mean per-session PnL into Greek contributions, second-order cross terms, and a residual, with mechanically similar patterns across models.The residual is negative for every model, at −7 to −32 dollars per session, and largely reflects the discontinuity of expiry settlement that a first-order expansion cannot reproduce.Source: Table 21 · p. 42, §H.1.2 · p. 43
  • Exact repricing accounts for 80% to 93% of absolute mark change at the step level and 76% to 88% after signed session aggregation.Spot contributes 0.71 to 0.77 of attributed magnitude, volatility 0.12 to 0.14, and time 0.03 to 0.04.Source: §H.1.2 · pp. 43–44, Table 22 · p. 44
  • Earnings-track option marks refresh in only 8% to 9% of minute steps, leaving step-level attribution limited to descriptive carried exposure.In 83% of Earnings step pairs the mark is unchanged while the underlying moves; exact repricing does not cure this measurement limit.Source: §H.1.2 · pp. 43–44

H.1.3 Earnings Bet

The Earnings Bet analysis links outcomes to the positions agents carried into announcements, their trading profiles, and the exposure to volatility contraction. It also examines early exit valuations and cautions against treating first-order Greek attribution as a full profit explanation.

Case studies

  • In one META example, a long strangle lost 18.97% despite a rise from $540 to $580 as implied volatility fell from a 125% peak to about 57%.The report attributes the loss to volatility collapse outpacing the underlying move.Source: §H.1.3 · p. 45, Figure 10 · p. 46
  • Across six illustrated cases, outcomes differed with the position held at announcement, not simply whether the structure was net-debit.The examples include gains when upside exposure met a rise, and losses when positions lacked the needed exposure or were already losing before the report.Source: §H.1.3 · p. 45, Figure 11 · p. 47

Exit policies

  • At 09:35 ET, valuation exceeded end-of-case return by a mean 1.60 and median 0.81 percentage points in 65.9% of the 82 cases with open positions.Most cases were already flat at that policy time; the comparison fixes decisions and excludes additional liquidation costs, so it does not establish an optimal exit rule.Source: §H.1.3 · p. 45, Table 23 · p. 45

Trading profile

  • Vertical and two-leg positions made up 60.8% to 93.0% of classified opening-intent groups across models, while call-leg shares ranged from 18.3% to 76.0%.These are opening-intent groups, not necessarily executed multi-leg fills, and their denominator differs from the per-case structure table.Source: §H.1.3 · pp. 45–47, Table 24 · p. 46
  • Average premium per case ranged from 20.9k to 35.2k dollars, with entry maturity of 3.6 to 4.1 days and holding periods of 1.5 to 2.3 days.Neither structure choice nor sizing ordered model outcomes in this panel.Source: Table 24 · p. 46, §H.1.3 · p. 47, Table 25 · p. 48

Structure and performance

  • 645 of 755 cases (85.4%) used net-debit structures, while credit spreads and naked short legs together accounted for 76 cases (10.1%).The prevalent long-premium positioning is exposed to the scheduled volatility contraction discussed in the attribution analysis.Source: §H.1.3 · p. 51, Table 26 · p. 52
  • The bear put spread was most common at 320 cases and averaged −0.85% across models; straddles and strangles averaged −4.19% and −4.52%, respectively.Bull call spreads averaged +0.84%; remaining structures had samples too small to support a ranking.Source: §H.1.3 · p. 52, Table 26 · p. 52

Greek and PnL attribution

  • Theta was −625 to −3,108 dollars per case for every model, while vega was +11 to +658 dollars; the residual ranged from −31,340 to +1,685 dollars per case.Because earnings combine large spot jumps with implied-volatility collapse, the first-order attribution describes carried exposure rather than explaining profits.Source: §H.1.3 · p. 52, Table 27 · p. 53

H.2 Ablation and Pairwise Detail

The appendix tests an execution guard and reports detailed context-ablation and model-pair comparisons. The guard’s measured mechanical effect was small, and only one earnings model pair remained significant after correction.

H.2 Ablation and Pairwise Detail

  • Only MiniMax-M3 versus GPT-5.6-Terra in Earnings was significant after Holm correction (p = 0.046); no 0DTE pair was significant.The appendix reports all 28 model pairs per episodic task.Source: §H.2 · p. 54, Table 30 · p. 56
  • Blocking duplicate closing intents changed mean returns by at most 0.30 percentage points, and no negative model mean became positive.The fixed-decision replay left the worst episode unchanged in 15 of 16 model–task cells; Qwen3-235B-A22B in 0DTE improved from −7.4% to −4.6%.Source: §H.2 · p. 54, Table 28 · p. 54
  • The context-ablation tables cover each tested profile–task–model cell, with paired return changes, confidence intervals, median changes, and adjusted p-values reported.Overlay cells have only two paired cases.Source: Table 29 · p. 55

H.3 Token Usage in Evaluation

The appendix reports evaluation token consumption and provides visual records of overlay returns, trading activity, and selected intraday sessions. These materials show how figures encode benchmark-relative paths and agent actions across scenarios.

H.3 Token Usage in Evaluation

  • The eight-model panel used 888.54 million tokens, or 89.67 million to 127.63 million per model, with input tokens dominating output.0DTE was the largest footprint for every model, at 60.34–93.53 million tokens over 102 sessions.Source: §H.3 · p. 57, Table 31 · p. 58

H.4.1 Portfolio Return for Overlay Scenarios

  • Figure 12 plots daily active return against matched buy-and-hold across five assets and both overlay mandates; flat segments indicate no option position.The benchmark holds each run’s own share count from the first close and leaves remaining cash idle.Source: Figure 12 · p. 59, §H.4.1 · p. 59

H.4.2 Trading Behavior Across Mandates

  • Figures 13–17 show cumulative returns and option-trade activity for five underlyings across six agents and the two mandates.The panels mark executed option trades and compare each path with a matched buy-and-hold line.Source: §H.4.2 · pp. 59–60, Figure 13 · p. 60, Figure 14 · p. 61

H.4.3 Samples for Intraday Scenarios

  • Figures 18–19 provide two 0DTE session samples each for DeepSeek-V4-Flash and GPT-OSS-120B, with actions and option positions shown alongside SPY paths.The sessions fall on either side of the 2025 daylight-saving change, allowing the Eastern session clock to be checked in both regimes.Source: §H.4.3 · p. 64, Figure 18 · p. 65, Figure 19 · p. 66

I Execution Quality and Quoted Liquidity in 0DTE

This section assesses 0DTE execution assumptions by comparing simulated fills with quoted spreads and observed liquidity, and reports methodological and research-practice disclosures, including LLM use and the absence of new assets, crowdsourcing, or human-subject experiments.

I Execution Quality and Quoted Liquidity in 0DTE

  • After half-spread charges, all three positive reported model means lose positivity, while the pooled mean falls from −0.199 to −0.417 percentage points without reliable evidence of a negative mean.Under a full spread, all eight models are negative; no individual model is significant at 5% after Holm correction.Source: Table 32 · p. 67, I Execution Quality and Quoted… · p. 68
  • The simulator’s slippage is economically negligible: 73.4% of fills incur zero slippage, and the $0.0001 charge when applied is about 50 times below the $0.0050 median quoted half-spread.The applied slippage is correctly signed, but reported returns should be read as excluding spread cost.Source: I Execution Quality and Quoted… · p. 67
  • Touch clustering is a tick-granularity artifact, not evidence of adverse execution: 69.5% of quotes are one cent wide, making their half-cent mid impossible for whole-cent fills.This clustering reflects price increments rather than the simulator’s execution quality.Source: I Execution Quality and Quoted… · p. 67
  • Quoted liquidity supports a half-spread cost bound: median spread is $0.010 (1.49% of mid), median near-touch depth is 116 contracts, and displayed size covers orders in 99.75% of fills.The median order is one contract and the largest is 17, so market impact beyond the touch is not binding at these sizes.Source: I Execution Quality and Quoted… · p. 67
  • Spread costs correlate 0.84 with per-episode trading intensity across models, so zero-spread leaderboards partly rank turnover rather than performance.GPT-OSS-120B trades 33.6 times per session versus MiniMax-M3’s 14.7 and drops from a share of first place on reported means to third after charging the spread.Source: I Execution Quality and Quoted… · p. 68
  • The analysis re-prices recorded 0DTE fills rather than modeling counterfactual decisions, and 15-minute quotes omit queue position, latency, and partial fills; equivalent Overlay and Earnings analysis is unavailable.Extending liquidity-aware evaluation requires quoted spread and depth in the agent’s observations; optional deterministic L1 replay is planned.Source: I Execution Quality and Quoted… · p. 68

10. Broader impacts

  • The paper answers N/A to releasing new assets.Source: §10 · p. 74

14. Crowdsourcing and research with human subjects

  • The paper answers N/A to crowdsourcing and human-subject research, stating that it does not include this kind of experiment.Source: §10 · p. 74

16. Declaration of LLM usage

  • The paper marks LLM usage as “Yes,” describing LLMs as core to benchmarking and evaluating performance in real-time investment.Source: §10 · p. 75

Related papers

Loading 2609.33470v2…