foxhunt/ml/tests/dqn_backtesting_integration_test.rs

//! WAVE 15 AGENT 34: DQN Backtesting Integration Test
//!
//! This test verifies that backtesting metrics are correctly integrated
//! into the DQN hyperopt objective function and that objectives vary
//! meaningfully across different hyperparameter configurations.
//!
//! ## Test Coverage
//!
//! 1. **Metrics Structure**: Verify DQNMetrics includes all 6 fields
//! 2. **Composite Objective**: Verify objective calculation formula
//! 3. **Objective Variance**: Verify objectives differ across trials
//! 4. **Backtesting Population**: Verify backtesting metrics are populated

use ml::hyperopt::adapters::dqn::{DQNMetrics, DQNParams, DQNTrainer};
use ml::hyperopt::traits::{HyperparameterOptimizable, ParameterSpace};

/// Test 1: Verify DQNMetrics structure includes backtesting fields
#[test]
fn test_dqn_metrics_structure() {
    let metrics = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 2.5,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: -50.0,
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: Some(1.5),       // Backtesting metric
        max_drawdown_pct: Some(-15.0), // Backtesting metric
        win_rate: Some(60.0),           // Backtesting metric
    };
    
    // Verify all fields are accessible
    assert_eq!(metrics.sharpe_ratio, Some(1.5));
    assert_eq!(metrics.max_drawdown_pct, Some(-15.0));
    assert_eq!(metrics.win_rate, Some(60.0));
    
    println!("✓ DQNMetrics structure includes all backtesting fields");
}

/// Test 2: Verify composite objective calculation
#[test]
fn test_composite_objective_calculation() {
    let metrics = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 5.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: 0.0,  // Neutral reward (maps to 0.5 score)
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: Some(2.0),       // 2.0/5.0 = 0.4 score
        max_drawdown_pct: Some(-10.0), // 10/100 = 0.1 penalty → 0.9 score
        win_rate: Some(60.0),           // 60/100 = 0.6 score
    };
    
    let objective = DQNTrainer::extract_objective(&metrics);
    
    // Expected calculation:
    // rl_reward_score = (0.0 + 10.0) / 20.0 = 0.5
    // sharpe_ratio_score = 2.0 / 5.0 = 0.4
    // drawdown_penalty = 10.0 / 100.0 = 0.1 → drawdown_score = 1.0 - 0.1 = 0.9
    // win_rate_score = 60.0 / 100.0 = 0.6
    //
    // composite_objective = 0.40 * 0.5 + 0.30 * 0.4 + 0.20 * 0.9 + 0.10 * 0.6
    //                     = 0.20     + 0.12     + 0.18     + 0.06
    //                     = 0.56
    //
    // Final objective = -0.56 (negated for minimization)
    let expected = -0.56;
    
    assert!(
        (objective - expected).abs() < 0.01,
        "Objective mismatch: expected {:.4}, got {:.4}",
        expected,
        objective
    );
    
    println!("✓ Composite objective calculation correct: {:.4}", objective);
}

/// Test 3: Verify objective varies across different configurations
#[test]
fn test_objective_variance_across_configs() {
    // Configuration 1: Good RL reward, poor backtesting
    let metrics1 = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 5.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: 5.0,  // High reward (maps to 0.75 score)
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: Some(0.5),       // Poor Sharpe (0.1 score)
        max_drawdown_pct: Some(-30.0), // High drawdown (0.7 score)
        win_rate: Some(45.0),           // Poor win rate (0.45 score)
    };
    
    // Configuration 2: Poor RL reward, good backtesting
    let metrics2 = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 5.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: -5.0, // Low reward (maps to 0.25 score)
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: Some(4.0),      // High Sharpe (0.8 score)
        max_drawdown_pct: Some(-5.0), // Low drawdown (0.95 score)
        win_rate: Some(75.0),          // High win rate (0.75 score)
    };
    
    // Configuration 3: Balanced performance
    let metrics3 = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 5.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: 0.0,  // Neutral reward (0.5 score)
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: Some(2.5),       // Medium Sharpe (0.5 score)
        max_drawdown_pct: Some(-15.0), // Medium drawdown (0.85 score)
        win_rate: Some(60.0),           // Medium win rate (0.6 score)
    };
    
    let obj1 = DQNTrainer::extract_objective(&metrics1);
    let obj2 = DQNTrainer::extract_objective(&metrics2);
    let obj3 = DQNTrainer::extract_objective(&metrics3);
    
    println!("Objective 1 (good RL, poor backtest): {:.4}", obj1);
    println!("Objective 2 (poor RL, good backtest): {:.4}", obj2);
    println!("Objective 3 (balanced): {:.4}", obj3);
    
    // Verify objectives are NOT identical
    assert_ne!(obj1, obj2, "Objectives should vary across configurations");
    assert_ne!(obj2, obj3, "Objectives should vary across configurations");
    assert_ne!(obj1, obj3, "Objectives should vary across configurations");
    
    // Calculate coefficient of variation (CV) to verify variance
    let mean = (obj1 + obj2 + obj3) / 3.0;
    let variance = ((obj1 - mean).powi(2) + (obj2 - mean).powi(2) + (obj3 - mean).powi(2)) / 3.0;
    let std_dev = variance.sqrt();
    let cv = (std_dev / mean.abs()) * 100.0;
    
    println!("Mean objective: {:.4}", mean);
    println!("Std dev: {:.4}", std_dev);
    println!("Coefficient of variation: {:.2}%", cv);
    
    // Verify reasonable variance (CV > 5%)
    assert!(
        cv > 5.0,
        "Coefficient of variation too low: {:.2}% (expected > 5%)",
        cv
    );
    
    println!("✓ Objectives vary meaningfully across configurations (CV={:.2}%)", cv);
}

/// Test 4: Verify backtesting metrics are populated (when available)
#[test]
fn test_backtesting_metrics_populated() {
    // Scenario 1: Backtesting metrics available
    let metrics_with_backtest = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 5.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: 0.0,
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: Some(2.0),
        max_drawdown_pct: Some(-15.0),
        win_rate: Some(60.0),
    };
    
    assert!(metrics_with_backtest.sharpe_ratio.is_some());
    assert!(metrics_with_backtest.max_drawdown_pct.is_some());
    assert!(metrics_with_backtest.win_rate.is_some());
    
    // Scenario 2: Backtesting metrics unavailable (None)
    let metrics_without_backtest = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 5.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: 0.0,
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: None,
        max_drawdown_pct: None,
        win_rate: None,
    };
    
    assert!(metrics_without_backtest.sharpe_ratio.is_none());
    assert!(metrics_without_backtest.max_drawdown_pct.is_none());
    assert!(metrics_without_backtest.win_rate.is_none());
    
    // Verify objectives differ between scenarios
    let obj_with = DQNTrainer::extract_objective(&metrics_with_backtest);
    let obj_without = DQNTrainer::extract_objective(&metrics_without_backtest);
    
    println!("Objective with backtesting: {:.4}", obj_with);
    println!("Objective without backtesting: {:.4}", obj_without);
    
    // With backtesting: uses actual metrics
    // Without backtesting: uses neutral fallback (0.5)
    assert_ne!(
        obj_with, obj_without,
        "Objectives should differ based on backtesting availability"
    );
    
    println!("✓ Backtesting metrics correctly handled (Some vs None)");
}

/// Test 5: Verify parameter space bounds (sanity check)
#[test]
fn test_parameter_space_consistency() {
    let bounds = DQNParams::continuous_bounds();
    
    // Verify we have 6 parameters (Wave 13 configuration)
    assert_eq!(bounds.len(), 6, "Expected 6 parameters in search space");
    
    // Verify all bounds are valid (lower < upper)
    for (i, (lower, upper)) in bounds.iter().enumerate() {
        assert!(
            lower < upper,
            "Invalid bounds for parameter {}: [{}, {}]",
            i, lower, upper
        );
    }
    
    println!("✓ Parameter space bounds are consistent");
}

/// Test 6: Verify objective normalization (prevents outliers)
#[test]
fn test_objective_normalization() {
    // Test outlier reward (should be clamped)
    let metrics_outlier = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 5.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: 100.0, // Extreme outlier (clamped to 10.0 → score 1.0)
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: Some(2.0),
        max_drawdown_pct: Some(-15.0),
        win_rate: Some(60.0),
    };
    
    let obj_outlier = DQNTrainer::extract_objective(&metrics_outlier);
    
    // Test normal reward (within expected range)
    let metrics_normal = DQNMetrics {
        train_loss: 0.5,
        val_loss: 0.4,
        avg_q_value: 5.0,
        final_epsilon: 0.01,
        epochs_completed: 100,
        avg_episode_reward: 10.0, // Max expected reward (score 1.0)
        buy_action_pct: 0.3,
        sell_action_pct: 0.3,
        hold_action_pct: 0.4,
        gradient_norm: 2.0,
        q_value_std: 1.5,
        sharpe_ratio: Some(2.0),
        max_drawdown_pct: Some(-15.0),
        win_rate: Some(60.0),
    };
    
    let obj_normal = DQNTrainer::extract_objective(&metrics_normal);
    
    // After clamping, both should have same objective (reward clamped to 1.0)
    assert!(
        (obj_outlier - obj_normal).abs() < 0.001,
        "Outlier should be clamped to same objective as max: outlier={:.4}, normal={:.4}",
        obj_outlier,
        obj_normal
    );
    
    println!("✓ Objective normalization prevents outlier domination");
}