Skip to main content

LLM Bullshit Detection: Tools and Approaches for Technical Analysis

· 3 min read
Max Kaido
Max Kaido
Architect

Following our observations about LLM behavior in technical analysis, a critical question emerges: How do we systematically detect when an LLM is confidently wrong or producing nonsensical analysis? This post explores the tools and approaches available for this challenge.

Available Tools and Approaches​

1. Automated Validation Tools​

1.1 Internal Consistency Checks​

  • Ranking System Invariants
    • Market order should not affect final ranking
    • Shuffling markets should produce consistent results
    • Transitive relationships must hold (if A > B and B > C, then A > C)
  • Numerical Validation
    • RSI thresholds (30/70) are well-defined and can be programmatically checked
    • Volume comparisons should maintain mathematical consistency
    • Price relationships must follow basic arithmetic rules

1.2 Cross-Model Validation​

  • Open Source Models
    • Different architectures might have different failure modes
    • Can use smaller, specialized models for specific validations
    • Cost-effective for initial screening
  • Commercial APIs
    • OpenAI/Anthropic as high-precision validators
    • Use sparingly due to cost constraints
    • Reserve for critical validations or tie-breaking

1.3 Time-Series Validation​

  • Historical Consistency
    • Previous analyses of same market should not wildly fluctuate
    • Trend changes should correlate with significant market events
    • Volume profile interpretations should be temporally consistent
  • Technical Indicator Math
    • Automated validation of indicator calculations
    • Cross-checking indicator relationships
    • Detecting mathematically impossible claims

2. Systematic Validation Approaches​

2.1 Known-Answer Testing​

  • Golden Dataset
    • Curated set of clear-cut technical analysis cases
    • Well-documented market conditions with expert analysis
    • Regular validation against these known cases
  • Edge Cases
    • Extreme market conditions
    • Corner cases in indicator values
    • Unusual volume profiles

2.2 Probabilistic Validation​

  • Ensemble Methods
    • Multiple models analyzing same data
    • Confidence weighted by model reliability
    • Disagreement as a signal for deeper investigation
  • Statistical Bounds
    • Expected ranges for various metrics
    • Probability thresholds for extreme claims
    • Time-series based probability checks

2.3 Domain-Specific Rules​

  • Technical Analysis Laws
    • RSI bounds and interpretation rules
    • Volume-price relationships
    • Trend definition requirements
  • Market Mechanics
    • Liquidity implications
    • Order book physics
    • Trading hour effects

3. Human-in-the-Loop Tools​

(Used sparingly and strategically)

3.1 Expert Review Triggers​

  • Anomaly Detection
    • Unusual pattern combinations
    • Unexpected confidence scores
    • Novel market behavior
  • Strategic Sampling
    • Regular audit of high-impact decisions
    • Review of edge cases for learning
    • Validation of new patterns

3.2 Feedback Loops​

  • Error Cataloging
    • Systematic recording of detected errors
    • Pattern recognition in failure modes
    • Continuous refinement of detection rules
  • Model Retraining Signals
    • Identifying systematic errors
    • Collecting correction examples
    • Prioritizing improvement areas

4. Meta-Validation Tools​

4.1 Process Validation​

  • Decision Trees
    • Clear validation pathways
    • Documented decision points
    • Failure mode handling
  • Audit Trails
    • Complete reasoning chains
    • Model confidence tracking
    • Validation step logging

4.2 System Health Metrics​

  • Validation Coverage
    • Percentage of decisions validated
    • Types of checks applied
    • Validation depth metrics
  • Error Rates
    • False positive tracking
    • Miss rate monitoring
    • Confidence correlation

Questions to Explore​

  1. Tool Integration

    • How do we efficiently combine these tools?
    • What's the optimal validation sequence?
    • How do we handle tool conflicts?

  2. Resource Optimization

    • When to use expensive vs cheap validation?
    • How to minimize API costs?
    • What's the minimal effective validation set?

  3. Scalability Concerns

    • How does validation time affect real-time analysis?
    • Can we parallelize validation effectively?
    • What's the maintenance overhead?

Remember: This is an initial exploration of available tools. The next post will focus on how to combine these tools into an effective bullshit detection system.