# TFT Hyperparameter Tuning Configuration (Recommended)
# Optimized for RTX 3050 Ti (4GB VRAM) and time-series forecasting

# Global tuning settings
global:
  optimization_direction: maximize  # Maximize combined objective
  pruning_enabled: true
  median_pruner:
    n_startup_trials: 5      # No pruning for first 5 trials (establish baseline)
    n_warmup_steps: 30       # Wait 30 epochs before starting to prune
    interval_steps: 10       # Check for pruning every 10 epochs
  sampler: TPE  # Tree-structured Parzen Estimator (best for < 50 trials)

# TFT model-specific search space
models:
  TFT:
    # ============================================================================
    # Training Parameters
    # ============================================================================
    epochs:
      type: int
      low: 30             # Minimum for TFT convergence
      high: 100           # Maximum for overnight run
      step: 10
      description: "Training epochs - TFT needs 30+ for attention convergence"

    learning_rate:
      type: float
      low: 0.0001         # Conservative for time-series stability
      high: 0.001         # Max for avoiding gradient explosion
      log: true           # Log scale for better exploration
      description: "Learning rate - lower range than DQN/PPO for time-series"

    batch_size:
      type: categorical
      choices: [32, 64]   # Max 64 for 4GB VRAM (TFT is memory-intensive)
      description: "Batch size - limited by VRAM and attention memory requirements"

    # ============================================================================
    # Architecture Parameters
    # ============================================================================
    hidden_dim:
      type: categorical
      choices: [128, 256]  # Max 256 for 4GB VRAM (512 requires 8GB+)
      description: "Hidden dimension - d_model in transformer architecture"

    num_heads:
      type: categorical
      choices: [4, 8, 16]  # Standard attention head counts
      description: "Attention heads - must divide hidden_dim evenly (128: 4/8, 256: 4/8/16)"

    num_layers:
      type: int
      low: 2              # Minimum for deep learning
      high: 4             # Max for 4GB VRAM (8 requires 12GB+)
      step: 1
      description: "LSTM layers - deeper networks = more memory + longer training"

    dropout_rate:
      type: float
      low: 0.0            # No dropout (rely on early stopping)
      high: 0.3           # Max for preventing underfitting
      step: 0.05
      description: "Dropout rate - regularization for preventing overfitting"

    # ============================================================================
    # Time-Series Specific Parameters
    # ============================================================================
    lookback_window:
      type: categorical
      choices: [30, 60, 120]  # 30 min, 1 hour, 2 hours (1-minute bars)
      description: "Lookback window - historical context for forecasting"

    forecast_horizon:
      type: categorical
      choices: [5, 10, 20]     # 5, 10, 20 minute predictions
      description: "Forecast horizon - prediction steps ahead"

    # ============================================================================
    # Early Stopping Parameters
    # ============================================================================
    early_stopping_patience:
      type: int
      low: 10             # Minimum patience (conservative)
      high: 30            # Maximum patience (aggressive)
      step: 5
      description: "Early stopping patience - epochs without improvement"

    early_stopping_threshold:
      type: float
      low: 0.00001        # Very sensitive (stop on tiny improvements)
      high: 0.001         # Less sensitive (require significant improvements)
      log: true
      description: "Early stopping threshold - minimum validation loss improvement"

# ============================================================================
# Objective Function Configuration
# ============================================================================
objective:
  type: weighted_combination
  weights:
    quantile_loss: -0.6    # 60% weight on forecast accuracy (minimize)
    sharpe_ratio: 0.4      # 40% weight on trading performance (maximize)

  # Forecast accuracy metrics
  forecast_metrics:
    - quantile_loss       # Primary metric (quantile regression loss)
    - rmse                # Root mean squared error
    - mae                 # Mean absolute error
    - quantile_coverage   # % of actuals within [0.1, 0.9] quantiles

  # Trading performance metrics
  trading_metrics:
    - sharpe_ratio        # Risk-adjusted returns (annualized)
    - max_drawdown        # Maximum equity decline
    - win_rate            # % of profitable trades
    - trade_frequency     # Trades per 1,000 bars (target: 50-150)

# ============================================================================
# Trial Configuration
# ============================================================================
trials:
  n_trials: 30            # Total trials (overnight run: 8-10 hours)
  timeout_per_trial: 1200 # 20 minutes max per trial (safety limit)

  # Trial execution
  n_jobs: 1               # Sequential trials (GPU memory safety)
  gc_after_trial: true    # Garbage collect after each trial

  # Pruning strategy
  enable_intra_trial_pruning: false  # Disabled (gRPC returns only final metrics)
  enable_inter_trial_pruning: true   # Enabled (compare final Sharpe ratios)

# ============================================================================
# Data Configuration
# ============================================================================
data:
  # Training data
  primary_symbol: "6E.FUT"     # Euro FX futures (7,223 bars)
  train_split: 0.80            # 80% training (5,778 bars)
  val_split: 0.20              # 20% validation (1,445 bars)

  # Cross-validation symbols (test generalization)
  cross_validation_symbols:
    - "ES.FUT"                 # E-mini S&P 500
    - "NQ.FUT"                 # Nasdaq futures
    - "ZN.FUT"                 # 10-Year Treasury

  # Data preprocessing
  normalize: true              # Z-score normalization
  remove_outliers: true        # Remove > 3 sigma outliers
  handle_gaps: "interpolate"   # Interpolate missing bars

# ============================================================================
# GPU Configuration
# ============================================================================
gpu:
  device: "cuda:0"             # RTX 3050 Ti
  max_vram_gb: 3.5             # Leave 0.5GB for system
  enable_mixed_precision: false # Disabled (stability issues with attention)
  enable_gradient_checkpointing: false # Disabled (not supported in candle)

# ============================================================================
# Success Criteria
# ============================================================================
success_criteria:
  # Minimum viable hyperparameters
  minimum_viable:
    quantile_loss: 0.3         # Reasonable forecast accuracy
    sharpe_ratio: 1.0          # Positive risk-adjusted returns
    validation_gap: 0.2        # Max 20% train-val loss difference
    trade_frequency_min: 50    # Minimum 50 trades per 7,223 bars
    trade_frequency_max: 150   # Maximum 150 trades per 7,223 bars

  # Production-ready hyperparameters
  production_ready:
    quantile_loss: 0.2         # Strong forecast accuracy
    sharpe_ratio: 1.5          # Excellent risk-adjusted returns
    rmse: 0.01                 # 1% prediction error
    quantile_coverage: 0.85    # 85% of actuals within [0.1, 0.9] quantiles
    cross_symbol_sharpe_tolerance: 0.3  # Within 30% across symbols

  # Study success
  study_success:
    minimum_viable_trials: 3   # At least 3 trials meet minimum viable (10% rate)
    production_ready_trials: 1 # At least 1 trial meets production-ready (3% rate)
    improvement_over_baseline: 0.2  # 20% improvement over baseline

# ============================================================================
# Monitoring & Logging
# ============================================================================
monitoring:
  log_level: "INFO"
  log_every_n_epochs: 10       # Log metrics every 10 epochs
  save_best_checkpoint: true   # Save checkpoint with best validation loss
  save_final_checkpoint: true  # Save checkpoint at end of training

  # Metrics to track
  tracked_metrics:
    - train_loss
    - val_loss
    - quantile_loss
    - rmse
    - mae
    - sharpe_ratio
    - max_drawdown
    - win_rate
    - trade_frequency
    - gradient_norm
    - learning_rate
    - epoch_duration

# ============================================================================
# Notes & Recommendations
# ============================================================================
#
# 1. VRAM Constraints:
#    - Batch size 64 max (TFT attention scales O(n^2) with sequence length)
#    - Hidden dim 256 max (512 requires 8GB+ VRAM)
#    - Num layers 4 max (8 requires 12GB+ VRAM)
#
# 2. Trial Duration Estimates:
#    - 30 epochs: ~8-10 minutes per trial
#    - 100 epochs: ~15-20 minutes per trial
#    - 30 trials × 15 min = 7.5 hours (overnight run)
#
# 3. Pruning Strategy:
#    - MedianPruner starts at epoch 30 (n_warmup_steps)
#    - Checks every 10 epochs (interval_steps)
#    - Expected time savings: 30-40%
#
# 4. Objective Function:
#    - 60% weight on forecast accuracy (quantile loss)
#    - 40% weight on trading performance (Sharpe ratio)
#    - TFT is a forecaster first, trader second
#
# 5. Cross-Validation:
#    - Train on 6E.FUT (Euro FX)
#    - Test on ES.FUT, NQ.FUT, ZN.FUT
#    - RMSE should be within 20% across symbols
#    - Sharpe ratio should be within 30% across symbols
#
# 6. Expected Best Hyperparameters:
#    - learning_rate: 0.0003
#    - batch_size: 64
#    - hidden_dim: 256
#    - num_heads: 8
#    - num_layers: 3
#    - lookback_window: 60
#    - forecast_horizon: 10
#    - dropout_rate: 0.15
#    - early_stopping_patience: 20
#
# 7. Expected Performance:
#    - Quantile loss: 0.15-0.25
#    - RMSE: 0.008-0.012
#    - Sharpe ratio: 1.3-1.8
#    - Trade frequency: 80-120 trades per 7,223 bars