Paper #1: Workshop Submission (LLM-Finance 2025) - Due Oct 26

[iAmGiG]

Context

Conference: LLM-Finance 2025 - The 2nd IEEE International Workshop on Large Language Models for Finance
Part of: IEEE BigData 2025 (December 8-11, 2025, Macau, China)
Submission Deadline: October 26, 2025 (12 days from today)
Paper Type: Workshop paper (not full journal paper)

Advisor guidance: "Currently the experimental study is good for the workshop, more sophisticated experimental study with some future work can do a journal publication"

Today's date: October 14, 2025

---

Paper Title Exploration

Advisor's Suggested Title

"something like 'Learning the Invisible Hand: Evaluating Large Language Models' Ability to Infer Market Forces under Data Obfuscation'"

Alternative Title Options

Option 1: Advisor's Suggestion (Economic Metaphor)

• "Learning the Invisible Hand: Evaluating Large Language Models' Ability to Infer Market Forces under Data Obfuscation"
• Pros: Catchy, references Adam Smith, emphasizes "market forces"
• Cons: "Invisible hand" is metaphorical (less precise)

Option 2: GEX/Gamma Exposure Focus (Technical)

• "Detecting Latent Market Patterns: LLM-Based Gamma Exposure Analysis with Obfuscation Testing"
• Pros: Uses "GEX" (recognizable to finance audience), "latent patterns" (catchy)
• Cons: More jargon-heavy

Option 3: Constraint Detection (Methodology Focus)

• "Obfuscation Testing for LLM Structural Reasoning: Detecting Dealer Hedging Constraints in Options Markets"
• Pros: Clear methodology contribution, specific domain
• Cons: Less catchy, longer

Option 4: Hybrid (Market Forces + GEX)

• "Inferring Latent Market Forces: LLM Detection of Gamma Exposure Patterns under Data Obfuscation"
• Pros: Combines "latent patterns" + "GEX" + "market forces"
• Cons: Slightly generic

Option 5: WHO→WHOM→WHAT Framework

• "From GEX to Action: LLM-Based Detection of Dealer Hedging Constraints via Obfuscation Testing"
• Pros: Emphasizes causal mechanism (WHO forces WHOM)
• Cons: Less emphasis on "market forces" metaphor

Option 6: Pattern Discovery Angle

• "Discovering Mechanical Market Patterns: Obfuscation Testing for LLM Structural Understanding in Finance"
• Pros: Emphasizes discovery, "mechanical patterns" (our classification)
• Cons: Less specific to GEX/dealer hedging

Recommended Title (Balances All Elements)

Final Recommendation:

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

Rationale:

• ✅ "Latent market forces" (broad appeal, user preference)
• ✅ "Gamma exposure" (GEX terminology, technical credibility)
• ✅ "Obfuscation testing" (methodology contribution)
• ✅ "Evaluating LLM" (clear scope)
• ✅ Length: 14 words (reasonable for finance/AI venues)

---

Workshop Requirements (IEEE Format)

Estimated Requirements (based on standard IEEE BigData workshop guidelines):

• Format: IEEE 2-column format
• Page Limit: 6-8 pages (typical for workshops)
• Template: IEEE Conference template (available on Overleaf)
• Submission: PDF via conference submission system
• Review: Single-blind (authors visible to reviewers)

Topics of Interest (LLM-Finance workshop):

• Large language models for financial applications
• Financial market prediction and analysis
• Risk management and compliance
• Novel LLM architectures for finance

---

Paper Structure (Workshop Version)

Abstract (150-200 words)

• Problem: Can LLMs detect structural patterns vs. memorize training data?
• Method: Obfuscation testing framework (remove dates/tickers)
• Domain: Dealer hedging constraints in options markets (gamma exposure)
• Results: [Detection rates, accuracy, cross-pattern validation]
• Contribution: Novel methodology for validating LLM structural reasoning

1. Introduction (1 page)

• Market context: Options explosion (2020-2024), 0DTE growth
• Problem: LLMs in finance - reasoning vs. memorization?
• Challenge: How to test true understanding?
• Our approach: Obfuscation testing + gamma exposure patterns
• Contributions: (1) Validation methodology, (2) Empirical evidence

2. Background & Related Work (0.75 pages)

• Dealer hedging constraints (academic foundation)
• Gamma exposure (GEX) and market microstructure
• LLMs in finance (sentiment analysis, forecasting)
• Gap: No validation of structural reasoning

3. Methodology (1.5 pages)

• Pattern taxonomy (MECHANICAL, NARRATIVE)
• WHO→WHOM→WHAT framework
• Obfuscation technique (dates→"Day T+N", tickers→generic)
• GEX calculation (Black-Scholes gamma aggregation)
• Outcome verification (forward returns, realized volatility)

4. Experimental Setup (1 page)

• Dataset: SPY 2024 (Q1, Q3, Q4) - 181 trading days
• Patterns tested: gamma_positioning, stock_pinning, 0dte_hedging
• LLM: GPT-4o-mini (tool calling) + O3-mini (reasoning)
• Validation criteria: ≥60% detection, ≥30 samples, obfuscation pass

5. Results (1.5 pages)

• Detection rates (per pattern, per quarter)
• Predictive accuracy (rule-based verification)
• Key finding: Detection ≠ Profitability (structural vs. economic)
• Obfuscation test results (passes all patterns)

6. Discussion (0.75 pages)

• Why detection stays constant while profits vary
• Implications for LLM structural reasoning
• Limitations: Single asset (SPY), one year, three patterns

7. Future Work & Conclusion (0.5 pages)

• Extend to individual equities (Issue Extend Validation to Individual Equities (5-7 Tickers) #87)
• Multi-year validation (2023, 2025)
• Cross-domain applications (supply chain, healthcare)
• Conclusion: LLMs can detect structural constraints via obfuscation testing

References (0.5 pages)

• Academic papers: Avellaneda (2003), Frey (1997), Gao (2024), etc.
• SqueezeMetrics white papers (industry validation)
• LLM reasoning papers (chain-of-thought, structural understanding)

Total: 6-8 pages (fits workshop format)

---

Task Checklist for User

Setup Tasks (Days 1-2: Oct 14-15)

• Set up Overleaf project

  • Create new project: "LLM-Finance-2025-Paper"
  • Select template: "IEEE Conference Template" (search in Overleaf)
  • Invite advisor as collaborator (share link)
  • Create initial structure (sections 1-7 as above)

• Gather validation results from HPCC

  • Coordinate with HPCC-side CC to get Issue Verify Validation Results on HPCC Database #86 results
  • Download YAML files for Q1, Q3, Q4 2024 (SPY only for now)
  • Extract key metrics: detection rates, accuracy, alpha

• Finalize title

  • Review title options with advisor
  • Get advisor approval on final title
  • Update Overleaf document header

Writing Tasks (Days 3-8: Oct 16-21)

• Draft Abstract (Day 3)

  • Problem statement (1-2 sentences)
  • Methodology overview (2-3 sentences)
  • Key results (2-3 sentences, use actual numbers once available)
  • Contribution statement (1 sentence)

• Write Introduction (Days 3-4)

  • Market context (options explosion, dealer hedging)
  • Problem motivation (LLM reasoning vs. memorization)
  • Research question clearly stated
  • Contributions list (3-4 bullet points)

• Write Methodology (Days 5-6)

  • Obfuscation technique (with examples)
  • Pattern taxonomy diagram (MECHANICAL vs. NARRATIVE)
  • WHO→WHOM→WHAT framework explanation
  • Validation criteria (60% threshold, 30 samples, obfuscation pass)
  • Figure: System architecture flowchart (use docs/SYSTEM_FLOW_SIMPLE.md)

• Write Results (Day 7)

  • Table 1: Detection rates by pattern/quarter
  • Table 2: Predictive accuracy comparison
  • Table 3: Alpha vs. Detection (showing independence)
  • Discussion of key finding (detection constant, profits vary)

• Write Background, Discussion, Future Work (Day 8)

  • Background: Academic papers (dealer hedging constraints)
  • Discussion: Why this proves structural reasoning
  • Future work: Individual equities (Issue Extend Validation to Individual Equities (5-7 Tickers) #87), multi-year, cross-domain

Refinement Tasks (Days 9-11: Oct 22-24)

• Get advisor feedback (Day 9)

  • Share Overleaf draft with advisor
  • Schedule meeting to review draft
  • Incorporate feedback on content/framing

• Create figures and tables (Day 10)

  • Figure 1: Obfuscation example (before/after)
  • Figure 2: System architecture (from docs/SYSTEM_FLOW_SIMPLE.md)
  • Table 1: Detection rates (actual numbers from Issue Verify Validation Results on HPCC Database #86)
  • Table 2: Predictive accuracy
  • Table 3: Detection vs. Profitability (key evidence)

• Proofread and format (Day 11)

  • Check IEEE format compliance (2-column, font size, margins)
  • Verify all citations formatted correctly
  • Check page count (target: 6-8 pages)
  • Run spell check
  • Check figure quality (300 DPI minimum for IEEE)

Submission Tasks (Day 12: Oct 25-26)

• Final review (Oct 25 morning)

  • Read paper start to finish
  • Verify all tables/figures referenced in text
  • Confirm all claims have evidence
  • Check references complete

• Generate PDF (Oct 25 afternoon)

  • Download from Overleaf
  • Verify PDF looks correct (no formatting issues)
  • Check file size (< 10MB typical limit)

• Submit to workshop (Oct 26 - DEADLINE)

  • Find submission link (WikiCFP or conference website)
  • Create account if needed
  • Upload PDF
  • Fill out metadata (title, authors, abstract, keywords)
  • Submit before 11:59 PM (check timezone!)
  • Save confirmation email

---

Writing Guidelines (Workshop Paper)

Tone & Style

What to emphasize:

• Novel methodology (obfuscation testing)
• Empirical evidence (detection rates, accuracy)
• Structural reasoning (constraints vs. correlations)

What to de-emphasize:

• Trading profitability (mention but don't focus)
• Implementation details (keep high-level)
• Future extensions (brief mention only)

Writing style:

• Clear and concise (workshop papers are shorter)
• Focus on methodology contribution
• Use examples to illustrate obfuscation
• Connect to broader LLM reasoning research

Key Messages

1. Problem: Hard to distinguish LLM reasoning from memorization
2. Solution: Obfuscation testing removes temporal context
3. Domain: Financial markets (dealer hedging constraints)
4. Evidence: 100% detection with obfuscation across 181 days
5. Contribution: Validation framework applicable beyond finance

Figures to Include

Figure 1: Obfuscation Example

BEFORE (Normal):
Date: 2024-01-05
Symbol: SPY
Net GEX: -$5.2B
Spot: $552.10

AFTER (Obfuscated):
Date: Day T+0
Symbol: INDEX_1
Net GEX: -$5.2B
Spot: [normalized]

Figure 2: System Architecture (from docs/SYSTEM_FLOW_SIMPLE.md)

• Data → Obfuscation → LLM → Detection → Verification

Figure 3: Detection vs. Profitability (scatter plot)

• X-axis: Quarter (Q1, Q3, Q4)
• Y-axis: Detection rate (bars) + Net alpha (line)
• Shows detection constant, profits vary

Tables to Include

Table 1: Detection Rates by Pattern/Quarter

Pattern	Q1 2024	Q3 2024	Q4 2024	Average
gamma_positioning	[TBD]	[TBD]	[TBD]	[TBD]
stock_pinning	[TBD]	[TBD]	[TBD]	[TBD]
0dte_hedging	[TBD]	[TBD]	[TBD]	[TBD]

Table 2: Predictive Accuracy

Pattern	Accuracy	Sample Size	Obfuscation Pass
gamma_positioning	[TBD]	181	✓
stock_pinning	[TBD]	181	✓
0dte_hedging	[TBD]	181	✓

Table 3: Detection vs. Economic Outcome

Quarter	Detection Rate	Predictive Accuracy	Net Alpha
Q1 2024	[TBD]	[TBD]	[TBD]
Q3 2024	[TBD]	[TBD]	[TBD]
Q4 2024	[TBD]	[TBD]	[TBD]

(Shows detection constant while profitability varies - key evidence)

---

Keywords (for Submission)

Suggested keywords (5-7):

• Large language models
• Market microstructure
• Obfuscation testing
• Gamma exposure
• Structural reasoning
• Financial markets
• Pattern detection

---

Dependencies

Blocking:

• Issue Verify Validation Results on HPCC Database #86: SPY validation results needed for all tables/figures
• HPCC access to retrieve YAML reports

Non-blocking (can proceed without):

• Issue Extend Validation to Individual Equities (5-7 Tickers) #87 (individual equities) - mention in Future Work only

---

Success Criteria

✅ Paper submitted by October 26, 2025 (deadline)
✅ Advisor approval on draft before submission
✅ 6-8 pages in IEEE format
✅ All tables/figures included with actual data
✅ Title finalized and approved
✅ References complete and properly formatted
✅ Abstract clearly states contribution and results

---

Timeline Summary

Date	Day	Task
Oct 14-15	1-2	Setup Overleaf, gather results, finalize title
Oct 16-17	3-4	Write abstract + introduction
Oct 18-19	5-6	Write methodology
Oct 20	7	Write results
Oct 21	8	Write background, discussion, future work
Oct 22	9	Advisor feedback + revisions
Oct 23	10	Create figures and tables
Oct 24	11	Proofread and format
Oct 25	12	Final review + generate PDF
Oct 26	DEADLINE	Submit to workshop

12 days - tight but feasible if validation results (Issue #86) are available soon

---

Contact Information

Workshop: LLM-Finance 2025
Conference Website: https://conferences.cis.um.edu.mo/ieeebigdata2025/workshops.html
WikiCFP: http://www.wikicfp.com/cfp/servlet/event.showcfp?eventid=189780
Workshop Organizer: QUANZHI LI

---

Priority

CRITICAL - 12-day deadline for first paper submission

---

Estimated Effort: 12 days of writing + advisor feedback + validation results
Target Completion: October 26, 2025 (submission deadline)

[iAmGiG]

Advisor Feedback on Title & Framing

Alternative Title Suggestion

Advisor's preference:

"Validating Market Microstructure Constraints Through LLM Pattern Detection: Evidence from Options Markets"

Comparison:

Title Option	Strengths	Weaknesses
"Validating Market Microstructure Constraints..."	• Clear contribution (validation methodology) • Academic tone (market microstructure) • Evidence-focused	• Longer (12 words) • Less catchy • Doesn't mention obfuscation
"Inferring Latent Market Forces..."	• Catchy ("latent forces") • Mentions GEX explicitly • Highlights obfuscation	• "Latent forces" less precise • Longer (14 words)

Recommendation: Use advisor's title for workshop submission (clearer academic framing)

---

Key Framing: Profitability Variation as Strength

How to present in paper (advisor's suggestion):

"Detection rates remain perfect (100%) and accuracy high (87-98%) across all quarters, while economic outcomes vary from +70bps to -1bps. This independence between pattern detection and profitability demonstrates that our methodology identifies structural market mechanics rather than fitting to profitable anomalies."

Where to include:

1. Abstract: "...pattern detection remains consistent despite economic outcome variation..."
2. Results section: Full table showing detection vs. profitability
3. Discussion section: "Why this proves structural reasoning (not overfitting)"

Paper narrative shift:

• ❌ OLD: "Detection = 100%, accuracy = 96%, alpha = +21 to -1 bps"
• ✅ NEW: "Detection constant across regimes while profitability varies - proving structural not statistical"

Key evidence table:

Quarter	Detection Rate	Predictive Accuracy	Net Alpha	Interpretation
Q1 2024	100%	96%	+70 bps	Bull market + high vol
Q3 2024	100%	98%	+4 bps	Market efficiency increase?
Q4 2024	100%	98%	-1 bps	Low vol regime
Trend	CONSTANT	CONSTANT	DECLINING	Detection ≠ Profitability

This table becomes centerpiece evidence that we're detecting structural constraints, not fitting to profits.

---

Updated Section for Paper

Results Section - Subsection 5.2: Detection vs. Economic Outcome

Draft text:

"A critical validation of our methodology is the independence between pattern detection and economic profitability. As shown in Table 3, detection rates remain perfect (100%) and predictive accuracy high (87-98%) across all three quarters, while economic outcomes vary substantially from +70bps in Q1 to -1bps in Q4.

This independence demonstrates that our obfuscation testing identifies structural market mechanics rather than fitting to profitable anomalies. If our methodology were overfitting or cherry-picking, we would expect both detection and profitability to decline together. Instead, the LLM consistently recognizes dealer hedging constraints even when economic exploitation of these constraints becomes less profitable.

The profitability decline likely reflects changing market conditions (volatility regime shifts, increased market efficiency, 0DTE market structure evolution) rather than detection failure. This separation between structural understanding and economic outcome is precisely what distinguishes mechanical patterns from narrative patterns in our taxonomy."

---

Action items:

• Update title in Overleaf to advisor's suggestion
• Add Table 3 (Detection vs. Profitability) to Results section
• Include above framing in Discussion section
• Use exact quote in abstract/introduction

[iAmGiG]

Paper #1 Status Update - October 16, 2025

Current Status: PAUSED - Awaiting Advisor Guidance

Issue #90 Discovery: During Paper #1 preparation, we discovered prompt bias in our validation methodology. The LLM was shown regime labels ("NEGATIVE_GAMMA") which essentially revealed the answer.

Revalidation Complete

All 3 patterns revalidated with unbiased prompts (no regime labels, no pattern hints):

Pattern	Biased Prompt	Unbiased Prompt	Delta	Accuracy
gamma_positioning	100%	69.4%	-30.6%	92.5%
stock_pinning	100%	67.4%	-32.6%	90.4%
0dte_hedging	100%	77.7%	-22.3%	90.8%
Average	100%	71.5%	-28.5%	91.2%

Key Finding: All 3 patterns still pass 60% mechanical threshold with unbiased prompts.

Advisor Guidance Needed

We've sent comprehensive update to advisor (docs/advisor_update_prompt_bias_oct2025.md) asking for guidance on presentation strategy:

• Option A: Lead with unbiased 71% (conservative, more defensible)
• Option B: Present both equally (ablation study approach)
• Option C: Lead with biased 100% (acknowledge as limitation)

Our Recommendation: Option A (lead with unbiased 71% as main result)

Work Continuing (While Waiting)

These tasks work regardless of advisor's choice:

1. ✅ Updating paper outline with all 3 presentation scenarios
2. 🔄 Drafting methods section (obfuscation + prompt ablation)
3. 🔄 Creating comparison tables and figures

Timeline Impact

• Sent: Oct 16, 2025
• Read: Same day
• Expected: Advisor response within 1-2 days
• Target: Paper still achievable for Oct 26 deadline once decision received

Why This Strengthens the Paper

The 71% unbiased detection rate is more academically defensible than 100%:

• Avoids "too good to be true" criticism
• Proves structural pattern detection without label leakage
• Shows rigorous methodology (we caught and fixed the bias)
• Transparent reporting of limitations
• 91% accuracy proves patterns are real (predictions materialize)

Status: Awaiting advisor input to finalize presentation approach and continue drafting.

[iAmGiG]

Advisor Decision: Option A Selected ✅

Decision: Lead with unbiased results (71% detection) as main finding, present biased results (100%) as sensitivity analysis.

Primary Results Now Committed (c303438)

Full Year 2024 - Unbiased Prompts (no regime labels):

Pattern	Detection	Accuracy	Status
gamma_positioning	69.4%	92.5%	✅ MECHANICAL
stock_pinning	67.4%	90.4%	✅ MECHANICAL
0dte_hedging	77.7%	90.8%	✅ MECHANICAL
Average	71.5%	91.2%	✅ PASS

All patterns exceed 60% mechanical threshold.

Paper Structure (Option A)

Results Section:

• Table 1 (Primary): Unbiased prompt results (71.5% detection, 91.2% accuracy)
• Table 2 (Sensitivity): Biased prompt results (100% detection, 91-94% accuracy)
• Emphasis: 71% as conservative lower bound, robust to prompt bias

Key Messages:

1. 71% proves structural detection without label leakage
2. 91% accuracy shows patterns materialize (real, actionable)
3. More academically defensible than 100% alone
4. Transparent methodology (tested both configurations)

Files Ready for Paper

Primary Evidence:

• gamma_positioning_SPY_2024_unbiased.yaml (242 days)
• stock_pinning_SPY_2024_unbiased.yaml (242 days)
• 0dte_hedging_SPY_2024_unbiased.yaml (242 days)

Sensitivity Analysis:

• Q3/Q4 2024 biased results (100% detection, existing files)

Documentation:

• docs/papers/biased_vs_unbiased_comparison.md (detailed analysis)
• docs/advisor_update_prompt_bias_oct2025.md (full context)

Next Steps

Paper Drafting (Chat B):

1. ✅ Finalize outline for Option A structure
2. ✅ Draft methods section (obfuscation + prompt templates)
3. ✅ Draft results section (lead with unbiased 71%)
4. ✅ Create tables (unbiased primary, biased sensitivity)
5. 🔄 Draft abstract and introduction

Timeline: Still on track for Oct 26 deadline.

Status: UNBLOCKED - All evidence committed, advisor decision final, paper drafting can proceed.

[iAmGiG]

Status Update: FIGURES COMPLETE ✅

All 8 must-have figures generated (Oct 16, 2025):

• ✅ Text sections complete (Main Chat)
• ✅ Validation data complete (Full 2024)
• ✅ References verified (13 core papers)
• ✅ Figures complete (Issues Paper #1 Figures: Core Methodology Visualizations (Must-Have Set 1) #91-93 closed)

Next: Tables and formatting (due Oct 24)
First Draft Target: Oct 26, 2025

[iAmGiG]

Status Update: Oct 18, 2025

Completion Status: 92% Complete

✅ COMPLETED:

• Text: All 8 sections drafted (50 KB)
• Validation Data: Full 2024 unbiased + Q2 biased (726 tests)
• References: 13 core papers verified, BibTeX ready
• Tables: 3 main tables embedded + LaTeX formatting guide
• Figures: 8 core figures generated (25 PNG files total)
• Documentation: Comprehensive organization complete

🔄 IN PROGRESS:

• Figures 4, 6, 7, 8: Chat B rebuilding with actual YAML data (discovered hardcoded values)
• LaTeX Conversion: Ready to begin once figures finalized

📋 REMAINING (Oct 18-26):

1. Chat B figure rebuild (Oct 18-19)
2. LaTeX conversion (Oct 20-24)
3. Final polish (Oct 25)
4. Submit to advisor (Oct 26)

Timeline: Still on track for Oct 26 deadline

Documentation: See docs/papers/paper1/DOCUMENTATION_INDEX.md for complete status