Paper: arXiv 2512.17923
Authors: Christopher Regan, Ying Xie
Abstract
We introduce obfuscation testing, a novel methodology for validating whether large language models detect structural market patterns through causal reasoning rather than temporal association. Testing three dealer hedging constraint patterns (gamma positioning, stock pinning, 0DTE hedging) on 242 trading days (95.6% coverage) of S&P 500 options data, we find LLMs achieve 71.5% detection rate using unbiased prompts that provide only raw gamma exposure values without regime labels or temporal context. The WHO-WHOM-WHAT causal framework forces models to identify the economic actors (dealers), affected parties (directional traders), and structural mechanisms (forced hedging) underlying observed market dynamics. Critically, detection accuracy (91.2%) remains stable even as economic profitability varies quarterly, demonstrating that models identify structural constraints rather than profitable patterns. When prompted with regime labels, detection increases to 100%, but the 71.5% unbiased rate validates genuine pattern recognition. Our findings suggest LLMs possess emergent capabilities for detecting complex financial mechanisms through pure structural reasoning, with implications for systematic strategy development, risk management, and our understanding of how transformer architectures process financial market dynamics.
Complexity vs Empirical Score
- Math Complexity: 2.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper’s mathematical complexity is low, focusing on causal logic (WHO→WHOM→WHAT) and gamma exposure metrics without advanced stochastic calculus or complex formulas. Empirical rigor is high, evidenced by a large dataset (242 days, 95.6% coverage), robust out-of-sample testing (obfuscation), and measurable performance metrics (71.5% detection rate, 91.2% validation).
Research Flowchart
flowchart TD
A["Research Goal: Validate if LLMs detect structural market patterns<br>via causal reasoning vs. temporal association"] --> B["Method: Obfuscation Testing"]
B --> C["Data: S&P 500 Options<br>242 Trading Days, 95.6% Coverage"]
C --> D["Computational Process: WHO-WHOM-WHAT Framework<br>Unbiased Prompts (Raw Gamma Values Only)"]
D --> E{"Evaluation"}
E --> F["Key Findings"]
F --> F1["LLM Detection Rate: 71.5%<br>(Unbiased context)"]
F --> F2["Detection Accuracy: 91.2%<br>(Stable across economic regimes)"]
F --> F3["Context Effect: 100%<br>(With regime labels)"]
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 novel obfuscation testing methodology to validate whether large language models detect structural market patterns through causal reasoning.
Source sentence
We introduce obfuscation testing, a novel methodology for validating whether large language models detect structural market patterns through causal reasoning rather than temporal association.
Assumes 5
- Practitioner gamma exposure calculations assume aggregate customers are net short calls and net long puts. market
Source sentence
Practitioner gamma exposure (GEX) calculations assume aggregate customers are net short calls and net long puts, an assumption that generally holds for index options.
- Dealers must hedge gamma exposure to maintain delta neutrality regardless of measurement methodology. market
Source sentence
Dealers must hedge gamma exposure to maintain delta neutrality regardless of measurement methodology, so detected patterns reflect fundamental market mechanics rather than artifacts of any particular estimation formula.
- Detection threshold is set at 60% confidence for mechanical classification. statistical
Source sentence
Detection threshold set at 60% confidence for MECHANICAL classification, with 95% confidence intervals using binomial proportion tests.
- Null hypothesis assumes 50% random detection; detection must significantly exceed this baseline. statistical
Source sentence
Our null hypothesis assumes 50% random detection; actual detection must significantly exceed this baseline (binomial test, p < 0.05).
Show 1 more
- Transaction costs are 5 basis points per trade. execution
Source sentence
TC is transaction costs (5 basis points per trade).
Data
- S&P 500 ETF (SPY) options data options public
Source sentence
We analyze S&P 500 ETF (SPY) options data covering 2024 from January 2 to December 31. Our dataset comprises 242 trading days (95.6% of 253 expected trading days), with end-of-day snapshots including complete bid/ask spreads, open interest, implied volatility, and Greeks for all available strikes and expirations.
Costs
Modeled
5 basis points per trade
Source sentence
TC is transaction costs (5 basis points per trade).
Finds 5
- LLMs achieve a 71.5% detection rate of dealer hedging patterns using unbiased prompts across 242 trading days. performance
Source sentence
we find LLMs achieve 71.5% detection rate using unbiased prompts that provide only raw gamma exposure values without regime labels or temporal context.
- Forward market behavior materializes for 91.2% of detections, indicating high predictive accuracy. predictive-accuracy
Source sentence
Result: 91.2% accuracy across all detections
- Causal attribution (WHO→WHOM→WHAT) is correct for 100% of detected patterns. statistical
Source sentence
Result: 100% correct causal attribution for detected patterns
- Detection rates remain stable across quarters (68.2%–73.6%) despite economic alpha declining from +21 bps to 0 bps. performance
Source sentence
Detection rates remain stable (100% to 84%) while net alpha declines from +21 bps to -1 bp across quarters, validating structural pattern detection independent of economic outcomes.
- Bootstrap resampling yields a mean detection rate of 71.3% (std 2.1%), confirming result stability. statistical
Source sentence
Bootstrap resampling (10,000 iterations) yields consistent detection rates (mean: 71.3%, std: 2.1%), confirming result stability.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 28 related papers). Coverage is partial, so this is neither replication nor refutation.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (43)
- The pricing of options and corporate liabilities (1973)
- Efficient capital markets: A review of theory and empirical work (1970)
- Market microstructure theory (1995)
- Empirical market microstructure (2007)
- Liquidity and market structure (1988)
- The impact of option hedging on the spot market volatility (2022) — link
- 0DTEs: Trading, Gamma Risk and Volatility Propagation (2025) — link
- Market Tremors: Quantifying Structural Risks in Modern Financial Markets (2021) — link
and 35 more
- Stock price clustering on option expiration dates (2005) — link
- Demand-based option pricing (2009) — link
- Asset fire sales (and purchases) in equity markets (2007)
- Is there price discovery in equity options? (2013)
- How markets slowly digest changes in supply and demand (2009)
- Market volatility and feedback effects from dynamic hedging (1997)
- Statistical arbitrage in the US equities market (2010)
- Ex-ante skewness and expected stock returns (2013) — link
- Why does the option to stock volume ratio predict stock returns? (2016)
- O/S: The relative trading activity in options and stock (2010)
- Language models are few-shot learners (2020)
- Chain-of-thought prompting elicits reasoning in large language models (2022)
- GPT-4 technical report (2023) — link
- Large language models are zero-shot reasoners (2022)
- Can ChatGPT forecast stock price movements? Return predictability and large language models — → 2304.07619
- FinGPT: Open-source financial large language models — → 2306.06031
- BloombergGPT: A large language model for finance (2023) — link
- Rebooting AI: Building artificial intelligence we can trust (2019)
- Beyond accuracy: Behavioral testing of NLP models with CheckList (2020) — link
- The adaptive markets hypothesis (2004)
- Gamma exposure and market dynamics (2019)
- The dark index: Measuring market-maker positioning (2020)
- Equity derivatives strategy: Gamma hedging flows and market impact (2017)
- The rise of zero days to expiration options: Market structure and implications (2023)
- 0DTE options: Market structure implications and intraday volatility (2023)
- Rule 15c3-1: Net capital rule for broker-dealers (2024)
- Rule 4210: Margin requirements (2024)
- Higher order G reeks (2007)
- Option markets and implied volatility: Past versus present (2009)
- The gamma trap (2013)
- Does option trading have a pervasive impact on underlying stock prices? (2022)
- Human Action: A Treatise on Economics (1949)
- Man, Economy, and State with Power and Market (1962)
- The SpotGamma Index: Measuring gamma exposure and market volatility (2021) — link
- Gamma exposure and the cross-section of stock returns (2017) — link
2 of 43 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.