foxhunt/docs/archive/agents/AGENT_72_HANDOFF.md

Agent 72 Handoff: DBN Parser Fix & Model Validation

From: Agent 71 (Model Validation Attempt) To: Agent 72 (DBN Parser Fix or Alternative Solution) Date: 2025-10-14 Priority: 🔴 CRITICAL - Blocks production model validation Estimated Time: 2-4 hours (Option A) or 30 minutes (Option B)

---

🎯 Your Mission

Primary Goal: Enable model validation by fixing DBN data loading (0 bars currently loaded)

Context: Agent 71 successfully fixed the backtest infrastructure to load trained DQN/PPO models, but DBN files return 0 OHLCV bars, blocking all validation work.

Choose ONE:

• Option A: Fix DBN parser (2-4 hours, permanent solution)
• Option B: Use synthetic data (30 min, temporary workaround)

---

📋 Option A: Fix DBN Parser (RECOMMENDED)

Current State

$ cargo run -p ml --example test_dbn_loading
✅ File loaded: 97KB
❌ OHLCV bars: 0  # SHOULD BE ~400-500
⚠️  Messages: 2 (type unknown)
⚠️  Warning: "Invalid message length: 0 at offset 23019"

Root Cause

• File: /home/jgrusewski/Work/foxhunt/data/providers/databento/dbn_parser.rs
• Issue: parse_batch() does not return ProcessedMessage::Ohlcv variants
• Known from Agent 63's work (Wave 160 Phase 3)

Your Tasks

Task 1: Debug parse_batch() (45-60 min)

Step 1: Add diagnostic logging

// In parse_batch()
for (i, msg) in messages.iter().enumerate() {
    debug!("Message {}: type={:?}, size={}", i, msg.rtype, msg.length);

    match msg.rtype {
        10 => { /* OHLCV */ },
        _ => warn!("Unexpected message type: {}", msg.rtype),
    }
}

Step 2: Check ProcessedMessage construction

// Verify OHLCV variant is being created
ProcessedMessage::Ohlcv {
    symbol: "6E.FUT".to_string(),
    timestamp: HardwareTimestamp::now(),
    open: Price::from_scaled_int(msg.open, 9),  // FIXED9
    high: Price::from_scaled_int(msg.high, 9),
    low: Price::from_scaled_int(msg.low, 9),
    close: Price::from_scaled_int(msg.close, 9),
    volume: Decimal::from_i64(msg.volume),
}

Step 3: Test with dbn-rs decode example

cd /tmp
cargo new dbn_test
cd dbn_test
cargo add dbn

# Create examples/decode.rs
cargo run --example decode /home/jgrusewski/Work/foxhunt/test_data/real/databento/ml_training_small/6E.FUT_ohlcv-1m_2024-01-02.dbn

Success Criteria:

• ✅ test_dbn_loading returns 400-500 OHLCV bars (not 0)
• ✅ No "Invalid message length" warnings
• ✅ comprehensive_model_backtest loads 400-500 bars per file

---

Task 2: Run Model Validation (30-45 min)

Once DBN parser fixed:

cargo run -p ml --example comprehensive_model_backtest --release

Expected Output:

🚀 COMPREHENSIVE ML MODEL BACKTESTING

Testing model: DQN on 6E.FUT
📊 Loading market data...
✅ Loaded 1,800 bars  # From 4 files × ~450 bars each
📈 PERFORMANCE METRICS
   Sharpe Ratio: 1.2
   Max Drawdown: 15.0%
   Win Rate: 55.0%
   Total PnL: $5,000

Testing model: PPO on 6E.FUT
📊 Loading market data...
✅ Loaded 1,800 bars
📈 PERFORMANCE METRICS
   Sharpe Ratio: 1.5
   Max Drawdown: 12.0%
   Win Rate: 58.0%
   Total PnL: $8,000

📊 SUMMARY
🏆 Best Model: PPO (Sharpe: 1.5)

---

Task 3: Create Validation Report (30-45 min)

File: /home/jgrusewski/Work/foxhunt/AGENT_72_MODEL_VALIDATION_REPORT.md

Template:

# Model Validation Report

## Executive Summary
- ✅/❌ DQN: PASS/FAIL (Sharpe: X.X, Drawdown: XX%)
- ✅/❌ PPO: PASS/FAIL (Sharpe: X.X, Drawdown: XX%)

## Validation Criteria
- PASS: Sharpe > 1.0 AND Drawdown < 20% AND Win Rate > 50%
- FAIL: Any metric below threshold

## DQN Results
| Metric | Value | Target | Status |
|--------|-------|--------|--------|
| Sharpe Ratio | X.X | > 1.0 | ✅/❌ |
| Max Drawdown | XX% | < 20% | ✅/❌ |
| Win Rate | XX% | > 50% | ✅/❌ |

## PPO Results
[Same table]

## Production Recommendation
- **Deploy DQN**: YES/NO
- **Deploy PPO**: YES/NO
- **Rationale**: [1-2 sentences]

## Next Steps
1. [If PASS] Paper trading integration
2. [If FAIL] Hyperparameter tuning

---

📋 Option B: Synthetic Data Workaround (FAST)

If DBN parser fix takes >2 hours:

Task 1: Generate Synthetic Data (15 min)

File: /home/jgrusewski/Work/foxhunt/ml/examples/comprehensive_model_backtest.rs

Replace load_market_data() with:

fn load_market_data_synthetic(symbol: &str, bars: usize) -> Result<Vec<MarketBar>> {
    println!("⚠️  Using SYNTHETIC data (DBN parser blocked)");

    let mut market_bars = Vec::new();
    let start_date = Utc::now() - chrono::Duration::days(5);
    let base_price = 1.0800; // 6E.FUT typical price

    for i in 0..bars {
        let timestamp = start_date + chrono::Duration::minutes(i as i64 * 5);

        // Realistic price movement with trend + noise
        let trend = (i as f64 / 100.0).sin() * 0.0050;
        let noise = ((i as f64 * 7.3).sin() * 0.0010) +
                    ((i as f64 * 13.7).cos() * 0.0005);
        let close = base_price + trend + noise;

        market_bars.push(MarketBar {
            timestamp,
            open: close - 0.0002,
            high: close + 0.0003,
            low: close - 0.0003,
            close,
            volume: 1000.0 + (i as f64 * 10.0).sin().abs() * 500.0,
        });
    }

    Ok(market_bars)
}

Task 2: Run Validation with Synthetic Data (10 min)

cargo run -p ml --example comprehensive_model_backtest --release

Document Limitations:

• ⚠️ Results use SYNTHETIC data (not real market data)
• ⚠️ Metrics are indicative only
• ⚠️ Real data validation still required before production

Task 3: Brief Report (5 min)

Note: Models validated with synthetic data, real validation pending DBN fix

---

🔍 Investigation Resources

Files to Check

1. DBN Parser: /home/jgrusewski/Work/foxhunt/data/providers/databento/dbn_parser.rs
2. Test Example: /home/jgrusewski/Work/foxhunt/ml/examples/test_dbn_loading.rs
3. Agent 63 Report: /home/jgrusewski/Work/foxhunt/AGENT_63_DBN_PARSER_FIX.md
4. Backtest Script: /home/jgrusewski/Work/foxhunt/ml/examples/comprehensive_model_backtest.rs

Test Data

• Location: /home/jgrusewski/Work/foxhunt/test_data/real/databento/ml_training_small/
• Files: 6E.FUT_ohlcv-1m_2024-01-0[2-5].dbn (4 files, 400KB total)
• Expected: ~400-500 OHLCV bars per file

Model Checkpoints

• DQN: /home/jgrusewski/Work/foxhunt/ml/trained_models/production/dqn_real_data/dqn_final_epoch500.safetensors (74KB)
• PPO: /home/jgrusewski/Work/foxhunt/ml/trained_models/production/ppo_real_data/ppo_actor_epoch_500.safetensors (42KB)

Commands

# Test DBN loading
cargo run -p ml --example test_dbn_loading --release

# Run backtest (after fix)
cargo run -p ml --example comprehensive_model_backtest --release

# Check results
ls -lh /home/jgrusewski/Work/foxhunt/results/backtest_results_*.json

---

🎯 Success Criteria

Option A Success (DBN Parser Fix)

• ✅ test_dbn_loading shows 400-500 OHLCV bars (not 0)
• ✅ comprehensive_model_backtest loads real data successfully
• ✅ Backtest generates performance metrics for DQN and PPO
• ✅ Validation report created with PASS/FAIL recommendations

Option B Success (Synthetic Data)

• ✅ Backtest runs with 1,800 synthetic bars
• ✅ Performance metrics generated
• ✅ Report notes limitations (synthetic data)
• ⚠️ Real validation still needed

---

🚨 Critical Notes

1. Don't Skip Validation: Models CANNOT go to production without validation
2. Real Data Preferred: Option A (DBN fix) is strongly recommended
3. Agent 63 Context: DBN parser was supposed to be fixed in Wave 160 Phase 3
4. Time Budget: If you have 2+ hours, choose Option A; if <2 hours, choose Option B

---

📞 Quick Start

Recommended Path (if you have 2-4 hours):

# 1. Verify the problem
cargo run -p ml --example test_dbn_loading --release
# Expected: 0 OHLCV bars (currently broken)

# 2. Add debug logging to parse_batch()
vim data/src/providers/databento/dbn_parser.rs

# 3. Test with real DBN decoder
cargo run --example decode_dbn_file test_data/real/databento/ml_training_small/6E.FUT_ohlcv-1m_2024-01-02.dbn

# 4. Fix ProcessedMessage::Ohlcv creation
# 5. Verify fix
cargo run -p ml --example test_dbn_loading --release
# Expected: 400-500 OHLCV bars

# 6. Run validation
cargo run -p ml --example comprehensive_model_backtest --release

# 7. Create report
vim AGENT_72_MODEL_VALIDATION_REPORT.md

Fast Path (if you have <2 hours):

# 1. Add synthetic data function
vim ml/examples/comprehensive_model_backtest.rs

# 2. Run backtest
cargo run -p ml --example comprehensive_model_backtest --release

# 3. Document limitations
vim AGENT_72_SYNTHETIC_VALIDATION_REPORT.md

---

📊 Expected Timeline

Option A (DBN Parser Fix)

• Task 1 (Debug): 45-60 min
• Task 2 (Validation): 30-45 min
• Task 3 (Report): 30-45 min
• Total: 2-4 hours

Option B (Synthetic Data)

• Task 1 (Generate): 15 min
• Task 2 (Run): 10 min
• Task 3 (Report): 5 min
• Total: 30 minutes

---

🏆 Final Deliverable

Option A:

• ✅ Fixed DBN parser (permanent solution)
• ✅ Real data validation complete
• ✅ Production deployment recommendation
• ✅ AGENT_72_MODEL_VALIDATION_REPORT.md

Option B:

• ⚠️ Temporary synthetic data validation
• ⚠️ Real validation still needed
• ⚠️ AGENT_72_SYNTHETIC_VALIDATION_REPORT.md

---

Good luck! Choose the path that fits your time budget. Option A is strongly preferred for production readiness.