foxhunt/BLOCKER_01_INVESTIGATION_REPORT.md

BLOCKER 1 Investigation Report: MLFeatureExtractor Analysis

Date: 2025-10-19 Investigator: Agent using Zen MCP + Task Tool Status: Investigation Complete Verdict: MLFeatureExtractor is NOT obsolete - needs careful update

---

Executive Summary

VERDICT: Option B - Careful Update Required

common::MLFeatureExtractor is NOT obsolete and is actively used in production paths. However, it is critically outdated and extracting only 30 features instead of 225. The ml::features::extraction module serves a different purpose (training-time batch feature extraction) while MLFeatureExtractor serves inference-time streaming feature extraction in production trading.

Critical Finding: This is a HIGH-RISK BLOCKER affecting live trading decisions. All 5 ML models are receiving incomplete feature vectors (30/225 = 13.3% completeness), potentially causing severely degraded predictions.

---

Comparison: MLFeatureExtractor vs ml::features::extraction

Aspect	common::MLFeatureExtractor	ml::features::extraction
Purpose	Inference-time streaming (online)	Training-time batch processing (offline)
Input	Single price/volume/timestamp	Array of OHLCV bars
Output	Vec<f64> (variable length)	Vec<[f64; 256]> (fixed 256-dim)
State	Stateful (maintains rolling windows)	Stateless (processes entire bar array)
Features	30 (Wave A + 4 Wave C)	256 (full feature set)
Usage	Production trading (real-time)	Model training (batch)
Location	common/src/ml_strategy.rs	ml/src/features/extraction.rs
Dependencies	None (self-contained)	Requires 50+ bars warmup
Architecture	Streaming feature extraction	Batch feature extraction

Key Difference: These are NOT interchangeable. They serve different architectural purposes.

---

Production Usage Confirmed

File: common/src/ml_strategy.rs:1423

pub fn new(lookback_periods: usize, min_confidence_threshold: f64) -> Self {
    Self {
        models: Arc::new(RwLock::new(models)),
        feature_extractor: Arc::new(RwLock::new(MLFeatureExtractor::new_wave_d(lookback_periods))),  // ← PRODUCTION USE
        model_performance: Arc::new(RwLock::new(HashMap::new())),
        min_confidence_threshold,
    }
}

Production Call Sites:

1. Trading Agent Service → AssetSelector → MLFeatureExtractor (assets.rs:136)
2. Trading Service → SharedMLStrategy → MLFeatureExtractor (ml_strategy.rs:1423)

---

Current vs Expected State

Current State (30 features):

• Wave A: 26 features (5 OHLCV + 21 technical)
• Wave C: 4 features (OBV Momentum, Volume Oscillator, A/D Line, EMA Ratio)
• Total: 30 features

Expected State (225 features):

• Wave A: 26 features
• Wave C Initial: 4 features
• Wave C Advanced: 175 features (3 microstructure + 10 alternative bars + 162 fractional diff)
• Wave D: 24 features (10 CUSUM + 5 ADX + 5 Transition Probs + 4 Adaptive Metrics)
• Total: 229 features (or 225 if we optimize)

Missing: 195 features (86.7% gap)

---

Risk Assessment

What Breaks If We Change It?

1. Model Dimension Mismatch:

   • All trained models expect 256 features (as per Wave D spec)
   • Current inference provides 30 features
   • Gap: 226 features (88% missing)
   • Impact: Models are either zero-padding (degraded accuracy) or throwing errors

2. Test Dependencies:

   • 31 tests in common/tests/ depend on 30-feature output
   • Tests explicitly assert: assert_eq!(features.len(), 30)
   • All tests currently passing (false security)

3. Production Services:

   • SharedMLStrategy used in Trading Service and Trading Agent Service
   • Change affects ALL live trading decisions

---

Recommendation: Safe Migration Path

Phase 1: Extend MLFeatureExtractor (2-3 hours)

Add Wave C Advanced Features (175 features):

• Microstructure (3)
• Alternative bars (10)
• Fractional differentiation (162)

Add Wave D Regime Features (24 features):

• CUSUM statistics (10)
• ADX directional (5)
• Transition probabilities (5)
• Adaptive metrics (4)

Phase 2: Update Model Adapters (1 hour)

Extend SimpleDQNAdapter to support 225 features:

pub fn with_feature_count(model_id: String, feature_count: usize) -> Self {
    let weights = match feature_count {
        26 => vec![0.02; 26],  // Wave A
        30 => vec![0.02; 30],  // Wave A + 4 Wave C
        36 => vec![0.02; 36],  // Wave B
        65 => vec![0.02; 65],  // Wave C partial
        225 => vec![0.01; 225],  // Wave D (NEW)
        _ => panic!("Unsupported feature count: {}", feature_count),
    };
    // ...
}

Phase 3: Test Migration (2 hours)

Update tests to expect 225 features:

#[test]
fn test_wave_d_feature_extraction() {
    let mut extractor = MLFeatureExtractor::new_wave_d(20);
    let features = extractor.extract_features(100.0, 1000.0, Utc::now());
    assert_eq!(features.len(), 225, "Wave D must extract 225 features");
}

Phase 4: Gradual Rollout (1 hour)

1. Keep legacy constructor (MLFeatureExtractor::new() → 30 features)
2. Use new constructor (MLFeatureExtractor::new_wave_d() → 225 features) in SharedMLStrategy
3. Monitor production prediction quality

---

Final Verdict

DO NOT REPLACE MLFeatureExtractor with ml::features::extraction

REASON: They serve fundamentally different purposes:

• MLFeatureExtractor: Streaming inference (real-time trading)
• ml::features::extraction: Batch training (offline model training)

CORRECT ACTION: Update MLFeatureExtractor to extract all 225 Wave D features while maintaining its streaming architecture.

ESTIMATED EFFORT: 6-8 hours total

• 2-3 hours: Implementation
• 2 hours: Testing
• 2-3 hours: Validation

PRIORITY: CRITICAL - This blocker prevents Wave D regime detection from functioning correctly in production.

---

Next Steps

Based on this investigation, we should proceed with:

1. Implementing Wave C advanced features in MLFeatureExtractor
2. Implementing Wave D regime features in MLFeatureExtractor
3. Updating model adapters to support 225 features
4. Updating tests to validate 225-feature extraction
5. Gradual rollout with production monitoring

DO NOT attempt to replace MLFeatureExtractor with ml::features::extraction - they are fundamentally incompatible.