GEX-LLM Pattern Analysis: LLM Structural Reasoning in Financial Markets

---

Overview

PhD research investigating whether Large Language Models can detect structural constraints in financial markets through genuine reasoning rather than training data memorization.

Core Innovation: Temporal obfuscation testing — stripping all dates, ticker symbols, and contextual markers from financial data, forcing LLMs to reason from numerical structure alone.

Application Domain: Options dealer gamma exposure (GEX) — the aggregate hedging constraints that market makers face due to their options inventory positions.

---

Key Results

Single-Day Detection (IEEE BigData 2025)

Metric	Result
Detection Rate	71.5% (obfuscated, unbiased prompts)
Predictive Accuracy	90.9% (forward returns)
Raw Chain Superiority	92.3% vs 61.5% GEX-assisted (+30.8pp)
Test Coverage	242 trading days (SPY, 2024)

Finding: LLMs reconstruct dealer positioning from raw strike-level data, outperforming pre-calculated metrics — empirical evidence that scalar GEX aggregation discards structural signal.

Multi-Day Regime Detection (AIAI 2026)

Metric	Result
2024 Detection	81.2% (persistent regimes)
2020 Detection	12.1% (pre-0DTE baseline)
Discrimination	69.1pp separation (φ = 0.672, p < 0.0001)
False Positives	0% on transitional/low-magnitude controls
Coverage	1,412 windows + 809 controls (2020–2025)

Finding: Detection tracks 0DTE options adoption — 3.7% (2021) → 100% (2024) — with GEX magnitude growing 360%, revealing a structural market reorganization.

Data note: the published 2024–2025 rates rest on inputs that lacked put-side GEX from 2023-09-29, and the 2024-vs-2020 baselines carried calendar dates. On the complete data the rule gives 10.3% (2024) and 13.6% (2025). Details: docs/papers/CORRECTIONS.md.

Detection ≠ Profitability

Stable detection (68–74% quarterly) persists while economic profitability collapses (Sharpe 1.8 → 0.1), confirming detected patterns are structural mechanics, not exploitable anomalies.

---

Research Papers

IEEE BigData 2025 ✅ Published — Single-Day Obfuscation Testing

Title: Inferring Latent Market Forces: Evaluating LLM Detection of Gamma Exposure Patterns via Obfuscation Testing

• arXiv: 2512.17923
• Venue: 2nd IEEE Workshop on LLMs for Finance @ IEEE BigData 2025 (Dec 2025, Macau)
• LaTeX: docs/papers/paper1/
• Headline: 71.5% obfuscated detection, 90.9% predictive accuracy on forward returns

AIAI 2026 ✅ Published — 30-Day Regime Detection

Title: Validating LLM Structural Reasoning: Detecting Persistent Market Regimes Through Temporal Obfuscation

• Venue: 22nd IFIP WG 12.5 International Conference on Artificial Intelligence Applications and Innovations (AIAI 2026), Springer, IFIP Advances in Information and Communication Technology, pp. 127–140
• DOI: 10.1007/978-3-032-30801-6_10
• LaTeX: docs/papers/paper2/aiai/
• Headline: 81.2% detection (2024) vs 12.1% (2020) — 69.1pp separation, φ = 0.672, p < 0.0001 over 2,221 evaluations

JRFM (MDPI) ✅ Published — Combined Methodology + Regime Detection

Title: Temporal Obfuscation Testing for LLM Structural Reasoning: From Single-Day Dealer Constraints to Persistent Market Regimes

A journal article that combines the obfuscation methodology validated at IEEE BigData 2025 with the multi-day regime-detection results from the AIAI 2026 paper, plus reviewer-driven additions: full prompt reproducibility, bootstrap CIs, χ² + Fisher contingency tests, threshold sensitivity, and a Markov-switching benchmark.

• Venue: Journal of Risk and Financial Management (MDPI)
• LaTeX: docs/papers/jrfm/
• Published: J. Risk Financial Manag. 2026, 19(6), 382, doi:10.3390/jrfm19060382

Electronics (MDPI) 🔄 In Preparation — Regime Re-Run on Complete Inputs

Title: Persistent Dealer-Gamma Regimes Under Temporal Obfuscation: A Re-Run on Complete Inputs with a Controlled LLM Re-Test

Extended version of the AIAI 2026 paper for the MDPI Electronics Special Issue "Timely AI Advancements and Interdisciplinary Synergies for Intelligent Modeling". It re-runs the regime test on complete daily GEX for SPY, QQQ and IWM (2008–2026) with gpt-6-luna, next to the published o4-mini labels, and stores every input, label and answer.

• LaTeX: docs/papers/electronics/
• Pipeline: scripts/validation/electronics/
• Status: co-author review; tracking issue #261

Future Directions

Research extensions not pursued within this repository (cross-asset generalization, intraday/per-strike analysis, GNN-based cross-asset hedging networks): docs/papers/extensions/.

---

Methodology

Obfuscation Testing

Raw:        SPY, 2024-03-15: Net GEX: -$32.9B, Flip: $485.00
Obfuscated: Day T+0, INDEX_1: Net GEX: -$32.9B, Flip: $485.00

Remove dates, tickers, events → preserve only quantitative structure → force structural reasoning.

WHO → WHOM → WHAT Causal Framework

Every detection must specify:

• WHO: The constrained actor (e.g., dealers with negative gamma)
• WHOM: The affected parties (e.g., directional traders)
• WHAT: The forced mechanism (e.g., pro-cyclical hedging amplifying volatility)

Regime Classification (30-Day Windows)

Criterion	Threshold	Purpose
Persistence	≥ 70% days same sign	Exceeds random binomial (~2.2σ)
Magnitude	≥ $5B average |GEX|	Economically significant positioning
Stability	≤ 5 sign flips	Sustained directional bias

---

Infrastructure

Database

• PostgreSQL 18.1: 81.8M contracts, 50 symbols, 2020–2025 (20.58 GB)
• Intraday snapshots: Yearly-partitioned table, 21 snapshots/day

Data Sources

• Alpha Vantage Premium: Historical options chains (1000 calls/min)
• Polygon.io: Stock price data (free tier)

LLM

• OpenAI o4-mini: Reasoning model via Batch API
• Cost: $11.07 for all 2,221 evaluations

---

Project Structure

gex-llm-patterns/
├── src/
│   ├── agents/              # LLM market mechanics agent
│   ├── analysis/            # Pattern library (15 patterns)
│   ├── gex/                 # GEX calculator (Black-Scholes)
│   ├── llm/                 # LLM integration
│   ├── validation/          # Obfuscation & regime classification
│   └── data_sources/        # Alpha Vantage, Polygon clients
├── scripts/
│   ├── validation/          # IEEE BigData + AIAI/JRFM validation pipelines, Electronics re-run
│   ├── analysis/            # Sensitivity analysis, figures
│   └── data_collection/     # Intraday OI monitor
├── docs/
│   ├── papers/
│   │   ├── paper1/          # Single-day obfuscation source — IEEE BigData 2025 (published)
│   │   ├── paper2/          # Regime detection source — AIAI 2026 (published)
│   │   ├── jrfm/            # Combined journal article — JRFM/MDPI (published 2026)
│   │   ├── electronics/     # Extension: regime re-run on complete inputs — MDPI Electronics (in preparation)
│   │   └── extensions/      # Forward-looking research directions (snapshot)
│   └── presentations/       # PhD symposium, fundamentals
├── reports/                  # Validation results (YAML)
└── config_defaults/          # Configuration templates

---

Getting Started

git clone https://github.com/iAmGiG/gex-llm-patterns.git
cd gex-llm-patterns
pip install -r requirements.txt

# Configure API keys
export OPENAI_API_KEY="your_key"
export POLYGON_API_KEY="your_key"

# Verify
python -c "from src.analysis.pattern_library import PatternLibrary; print('OK')"

Research Ethics

• Academic research only — not trading advice
• Public data only — Alpha Vantage options chains
• Open source — methodology and code fully available
• Not financial advice — detection rates do not imply profitable strategies

Citation

@inproceedings{regan2025obfuscation,
  author = {Regan, Christopher and Xie, Ying},
  title = {Inferring Latent Market Forces: Evaluating {LLM} Detection of
           Gamma Exposure Patterns via Obfuscation Testing},
  booktitle = {2nd IEEE International Workshop on Large Language Models
               for Finance, IEEE International Conference on Big Data},
  year = {2025},
  publisher = {IEEE},
  url = {https://arxiv.org/abs/2512.17923}
}

License

GNU Affero General Public License v3.0 — see LICENSE.

---

Last Updated: April 2026 Contact: Christopher Regan ([email protected]) · Ying Xie ([email protected]) Institution: Kennesaw State University, College of Computing and Software Engineering