Skip to content

Publications

Jump to bottom
WormsCanned edited this page Oct 26, 2025 · 1 revision

Publications & Presentations

Research outputs from the GEX LLM Patterns project.

Papers

Paper #1: Obfuscation Testing Baseline ✅

Title: "Validating Large Language Model Understanding of Market Microstructure Through Obfuscation Testing"

Status: ✅ Submitted (October 26, 2025)

Venue: LLM-Finance 2025 Workshop @ IEEE BigData 2025

Type: 4-6 page workshop paper

Abstract:

Large Language Models (LLMs) demonstrate impressive performance on financial tasks, but a fundamental question remains unanswered: do they truly understand market constraints, or simply memorize patterns from training data? We introduce obfuscation testing—a novel validation methodology that strips all temporal context (dates, tickers, events) and forces LLMs to reason purely from market structure.

Testing on options market dealer constraints across 242 trading days (2024), we find LLMs detect gamma hedging constraints with 71.5% detection rate and 91.2% predictive accuracy without any temporal information. Critically, detection remains stable (100%) even as economic profitability declines to zero (Q1→Q4 2024), proving LLMs detect structural constraints rather than profit opportunities.

Multi-pattern validation across three narrative framings (gamma positioning, stock pinning, 0DTE hedging) confirms LLMs identify the underlying causal mechanism, not surface keywords. Our framework provides rigorous validation that LLMs can reason about complex market dynamics without temporal context leakage.

Key Results:

  • Detection: 71.5% average (unbiased prompts, 726 tests)
  • Accuracy: 91.2% (predictions materialize)
  • Sample: 242 trading days × 3 patterns
  • Validation: Full 2024 (Q1, Q3, Q4)

Full Paper Content: docs/papers/paper1/

GitHub Issue: #88 (Paper #1 Status)

Paper #2: Sequential GEX Analysis 📋

Title (Proposed): "Temporal Dynamics of Dealer Constraints: Sequential Gamma Exposure Analysis with LLMs"

Status: 📋 Planned (Q1 2026)

Venue: Journal submission (6-8 pages)

Research Questions:

  1. Can LLMs detect constraint trajectories (not just snapshots)?
  2. Does sequential context improve predictive accuracy?
  3. What temporal patterns emerge in dealer hedging behavior?

Methodology:

  • 5-day lookback windows (Day T-4 to T+0)
  • Maintain obfuscation (no real dates)
  • New pattern taxonomy: Accumulation, relief, reversal, persistence

Expected Timeline:

  • Implementation: 5 days
  • Analysis/Writing: 2-3 weeks
  • Submission: Q1 2026

GitHub Issue: #89 (Sequential GEX Analysis)

Dependency: Paper #1 acceptance

Paper #3: Cross-Asset Generalization 📋

Title (Proposed): "Cross-Asset Validation of LLM Market Microstructure Understanding"

Status: 📋 Planned (Q2 2026)

Venue: Journal submission (8-10 pages)

Research Questions:

  1. Does obfuscation testing generalize beyond SPY index options?
  2. Do dealer constraints differ between index and single-name options?
  3. Can LLMs detect stock-specific vs market-wide patterns?

Methodology:

  • Test on 10-20 individual stocks (AAPL, MSFT, NVDA, TSLA, etc.)
  • Use sequential analysis if Paper #2 validates it
  • Compare dealer dynamics: Index vs single-name

Expected Timeline:

  • Data Collection: 1-2 weeks
  • Validation: 1 week
  • Analysis/Writing: 2-3 weeks
  • Submission: Q2 2026

GitHub Issue: #6 (Cross-asset validation)

Dependencies: Paper #1 acceptance, Paper #2 submission

Presentations

PhD Symposium 2025 ✅

Event: Academic PhD Research Symposium

Date: October 22, 2025

Title: "Testing LLM Structural Reasoning in Complex Systems: Options Market Case Study"

Type: 15-minute presentation + Q&A

Audience: PhD students, faculty, academic researchers

Materials:

Key Topics:

  1. Obfuscation testing methodology
  2. WHO→WHOM→WHAT framework
  3. Detection-profitability divergence finding
  4. Multi-pattern generalization evidence

Outcome: ✅ Successfully delivered

Documentation: docs/presentations/phd_symposium_2025.md

GitHub Issue: #95 (Presentation Diagrams)

IEEE BigData 2025 (Planned) 📅

Event: IEEE International Conference on Big Data 2025

Workshop: LLM-Finance 2025

Date: December 2025 (tentative)

Status: Paper submitted, awaiting acceptance

Presentation: If accepted, will present Paper #1 findings

Format: 20-minute talk + poster session

Conference Targets (Future)

Potential venues for Papers #2-3 and beyond:

Finance Conferences

  • AFA (American Finance Association)

    • Premier finance conference
    • Rigorous peer review
    • High-impact venue

  • WFA (Western Finance Association)

    • Strong finance focus
    • Academic audience

  • MFA (Midwest Finance Association)

    • Regional but well-regarded

Machine Learning Conferences

  • NeurIPS (Finance + ML track)

    • Top-tier ML conference
    • Emerging finance applications track

  • ICML (Finance + ML workshop)

    • International ML focus
    • Workshop on finance applications

Specialized Conferences

  • QWAFAFEW (Quantitative Work Alliance For Applied Finance, Education, and Wisdom)

    • Practitioner-focused
    • Options/derivatives emphasis

  • MFA (Market Microstructure Conference)

    • Specialized venue
    • Dealer behavior focus

Technical Reports

Multi-Pattern Validation 2024 ✅

Title: "Full Year Multi-Pattern Validation Results (2024)"

Date: October 2025

Location: docs/archive/multipattern_validation_2024.md

Contents:

  • Comprehensive analysis of 9 quarter-pattern combinations
  • Detection-profitability divergence discovery
  • Pattern consolidation findings
  • Alpha decline investigation

Status: ✅ Complete

Obfuscation Testing Framework Guide ✅

Title: "Obfuscation Testing Methodology: Complete Guide"

Date: October 2025

Location: docs/guides/obfuscation-testing-explained.md

Contents:

  • Methodology overview
  • Implementation details
  • Success criteria
  • Comparison to alternatives

Status: ✅ Complete

Pattern Taxonomy Framework ✅

Title: "Pattern Taxonomy and Validation Framework"

Date: October 2025

Location: docs/guides/validation-framework.md

Contents:

  • Mechanical vs narrative classification
  • WHO→WHOM→WHAT framework
  • Pattern testing process

Status: ✅ Complete

Datasets & Reproducibility

Validation Results (Public)

Location: reports/validation/pattern_taxonomy/

Files (9 YAML reports):

  • gamma_positioning_SPY_2024Q1.yaml
  • gamma_positioning_SPY_2024Q3.yaml
  • gamma_positioning_SPY_2024Q4.yaml
  • stock_pinning_SPY_2024Q1.yaml
  • stock_pinning_SPY_2024Q3.yaml
  • stock_pinning_SPY_2024Q4.yaml
  • 0dte_hedging_SPY_2024Q1.yaml
  • 0dte_hedging_SPY_2024Q3.yaml
  • 0dte_hedging_SPY_2024Q4.yaml

Format: Human-readable YAML with per-day results

Availability: ✅ Public (included in repository)

Code & Scripts (Public)

Validation Scripts: scripts/validation/

  • validate_pattern_taxonomy.py - Single pattern validation
  • validate_all_patterns.py - Batch validation

Core Components: src/

  • agents/market_mechanics_agent.py - LLM orchestration
  • gex/gex_calculator.py - Gamma exposure metrics
  • validation/outcome_calculator.py - Forward returns
  • data_sources/data_obfuscator.py - Obfuscation logic

Availability: ✅ Public (AGPL-3.0 license)

Media & Outreach

Blog Posts

Blog: Post Essentials

Topics (potential):

  1. "How to Know if LLMs Actually Understand Finance"
  2. "Obfuscation Testing: Preventing LLM Cheating"
  3. "Why Detection ≠ Profitability: Lessons from 2024"

Status: Planned post-Paper #1 acceptance

Talks & Workshops

Potential Invitations:

  • Academic seminars (finance/CS departments)
  • Industry talks (quant firms, exchanges)
  • Workshops on LLM validation methodologies

Status: Open to invitations

Collaborations

Open to Collaboration On

  1. Comparative LLM Testing:

    • Test Claude, o3-mini, open-source models
    • Compare reasoning capabilities

  2. Cross-Asset Extension:

    • Individual equity options data
    • Cross-asset validation (FX, rates, etc.)

  3. Real-Time Applications:

    • Production constraint monitoring
    • Market surveillance systems

  4. Formal Verification:

    • Mathematical proofs of constraints
    • Hybrid LLM + formal methods

Contact

GitHub Issues: github.com/iAmGiG/gex-llm-patterns/issues

Email: See profile at @iAmGiG

Research Roadmap: See RoadMap for future directions

Citation

If you use this work in your research, please cite:

@inproceedings{regan2025obfuscation,
  title={Validating Large Language Model Understanding of Market Microstructure Through Obfuscation Testing},
  author={Regan, Chris},
  booktitle={IEEE International Conference on Big Data (LLM-Finance Workshop)},
  year={2025},
  note={Submitted}
}

Repository: https://github.com/iAmGiG/gex-llm-patterns

License: AGPL-3.0

Last Updated: October 25, 2025

Clone this wiki locally