foxhunt/ml/tests/bug15_portfolio_compounding_test.rs

// Bug #15: Portfolio Reset Per Epoch
//
// Root Cause: ml/src/trainers/dqn.rs:1052 resets portfolio to $100k at START of every epoch
// Impact: Constant rewards (~0.004 ± 0.0001), zero learning signal, no compounding
//
// Expected Behavior: Portfolio should compound across entire training run
// - Epoch 1: $100k → $105k (reward = 0.05)
// - Epoch 2: $105k → $110k (reward = 0.05, compounded on higher base)
// - Epoch 100: $500k → $550k (reward = 0.50, 10x learning signal!)

use anyhow::Result;
use ml::trainers::dqn::{DQNHyperparameters, DQNTrainer};

/// Helper: Create minimal DQN trainer for testing
fn create_test_trainer() -> Result<DQNTrainer> {
    let mut hyperparams = DQNHyperparameters::conservative();

    // Override for fast test
    hyperparams.epochs = 3;
    hyperparams.batch_size = 32;
    hyperparams.buffer_size = 10000;
    hyperparams.min_replay_size = 100;
    hyperparams.checkpoint_frequency = 1;
    hyperparams.min_epochs_before_stopping = 1000;
    hyperparams.early_stopping_enabled = false;

    DQNTrainer::new(hyperparams)
}

/// Helper: Calculate variance of a Vec<f32>
fn calculate_variance(values: &[f32]) -> f32 {
    if values.len() < 2 {
        return 0.0;
    }
    let mean = values.iter().sum::<f32>() / values.len() as f32;
    let variance = values.iter()
        .map(|v| {
            let diff = v - mean;
            diff * diff
        })
        .sum::<f32>() / values.len() as f32;
    variance
}

#[test]
fn test_portfolio_reset_removed_from_training_loop() {
    // This test verifies that portfolio.reset() is NOT called in the training loop
    // We check this by examining the source code directly

    let source_code = include_str!("../src/trainers/dqn.rs");

    // Count occurrences of ACTUAL calls (not comments with "REMOVED:")
    // We look for lines that execute reset, not document it
    let reset_count = source_code
        .lines()
        .filter(|line| {
            let trimmed = line.trim();
            // Only count if:
            // 1. Contains "portfolio_tracker.reset()"
            // 2. NOT a comment (doesn't start with //)
            // 3. NOT inside a multiline comment
            trimmed.contains("portfolio_tracker.reset()")
                && !trimmed.starts_with("//")
                && !trimmed.starts_with("*")
                && !trimmed.contains("REMOVED:")
        })
        .count();

    // We expect ZERO occurrences of actual reset calls
    assert_eq!(reset_count, 0,
        "Found {} ACTUAL calls to portfolio_tracker.reset() in dqn.rs. \
        Bug #15 fix requires removing reset from training loop (line ~1052). \
        Portfolio should compound across all epochs, not reset per epoch. \
        Note: Comments with 'REMOVED:' are OK and expected.",
        reset_count);

    println!("✅ Test PASSED: No portfolio_tracker.reset() calls found in training loop");
    println!("   Portfolio will compound across all epochs as expected");
}

#[test]
fn test_portfolio_compounding_explanation() {
    // This test documents the expected behavior after Bug #15 fix

    println!("\n=== Bug #15 Fix: Portfolio Compounding ===");
    println!("");
    println!("BEFORE (Bug #15 - BROKEN):");
    println!("  Epoch 1: $100k → $105k (reward = +0.05)");
    println!("  Epoch 2: $100k → $104k (reward = +0.04) ← RESET TO $100k!");
    println!("  Epoch 3: $100k → $106k (reward = +0.06) ← RESET TO $100k!");
    println!("  Result: Constant rewards ~0.004 ± 0.0001 (ZERO learning signal)");
    println!("");
    println!("AFTER (Bug #15 - FIXED):");
    println!("  Epoch 1: $100k → $105k (reward = +0.05)");
    println!("  Epoch 2: $105k → $110k (reward = +0.05, 5% on higher base)");
    println!("  Epoch 3: $110k → $116k (reward = +0.06, compounds further)");
    println!("  Result: Increasing rewards with variance (STRONG learning signal)");
    println!("");
    println!("KEY INSIGHT:");
    println!("  By removing portfolio_tracker.reset() from line 1052,");
    println!("  the portfolio compounds across ALL epochs, creating");
    println!("  increasing reward variance that guides DQN learning.");
    println!("");

    // This test always passes - it's documentation
    assert!(true, "Portfolio compounding documented");
}