# Best Hyperparameters for ML Models
#
# These hyperparameters are the result of Optuna tuning studies
# or manual optimization. Update this file when better parameters
# are discovered through tuning or experimentation.
#
# Last Updated: 2025-10-14
# Tuning Method: Manual (defaults based on research)
# Next Tuning: 2025-01-15 (post-Q4 retraining)

# DQN (Deep Q-Network)
dqn:
  learning_rate: 0.0001
  batch_size: 128
  gamma: 0.99
  epsilon_start: 1.0
  epsilon_end: 0.01
  epsilon_decay: 0.995
  buffer_size: 100000
  epochs: 200
  checkpoint_frequency: 20
  early_stopping_enabled: true
  q_value_floor: 0.5
  min_loss_improvement_pct: 2.0
  plateau_window: 30
  min_epochs_before_stopping: 50
  # Tuning notes:
  # - Batch size 128 optimal for RTX 3050 Ti (4GB VRAM)
  # - Learning rate 0.0001 provides stable convergence
  # - Epsilon decay 0.995 balances exploration vs exploitation

# PPO (Proximal Policy Optimization)
ppo:
  learning_rate: 0.0003
  batch_size: 64
  gamma: 0.99
  gae_lambda: 0.95
  clip_epsilon: 0.2
  value_loss_coef: 0.5
  entropy_coef: 0.01
  epochs: 200
  checkpoint_frequency: 20
  early_stopping_enabled: true
  min_loss_improvement_pct: 1.5
  plateau_window: 30
  # Tuning notes:
  # - Batch size 64 to fit both actor and critic in GPU memory
  # - Clip epsilon 0.2 standard PPO value
  # - Entropy coef 0.01 encourages exploration

# MAMBA-2 (State Space Model)
mamba2:
  learning_rate: 0.0001
  batch_size: 32
  gamma: 0.99
  state_size: 16
  d_model: 128
  n_layers: 4
  expansion_factor: 2
  epochs: 150
  checkpoint_frequency: 15
  early_stopping_enabled: true
  min_loss_improvement_pct: 2.5
  plateau_window: 20
  # Tuning notes:
  # - Batch size 32 due to large model size (150-500MB)
  # - State size 16 captures sufficient market dynamics
  # - 4 layers provides good depth without overfitting

# TFT (Temporal Fusion Transformer)
tft:
  learning_rate: 0.001
  batch_size: 64
  gamma: 0.99
  hidden_size: 128
  num_heads: 4
  num_layers: 3
  dropout: 0.1
  epochs: 100
  checkpoint_frequency: 10
  early_stopping_enabled: true
  min_loss_improvement_pct: 2.0
  plateau_window: 15
  # Tuning notes:
  # - Higher learning rate 0.001 works well for TFT
  # - 4 attention heads balance computation vs capacity
  # - Dropout 0.1 prevents overfitting on financial data

# TLOB (Temporal Limit Order Book)
tlob:
  # TLOB is inference-only (rules-based engine)
  # No training hyperparameters needed
  inference_only: true
  features:
    - price_levels: 10
    - volume_imbalance: true
    - microstructure_features: 51
  notes: |
    TLOB model does not require training. It uses rules-based
    microstructure analytics on Level-2 order book data.
    Training would require tick-by-tick order book snapshots,
    which are not currently available in our data pipeline.

# LIQUID (Liquidity Neural Network)
liquid:
  learning_rate: 0.0005
  batch_size: 64
  gamma: 0.99
  hidden_size: 64
  num_liquid_layers: 2
  epochs: 120
  checkpoint_frequency: 12
  early_stopping_enabled: true
  # Tuning notes:
  # - Liquid networks are efficient, can use moderate batch size
  # - 2 liquid layers provide good temporal dynamics

# Global Training Settings
global:
  device: "cuda"  # Use GPU if available, fallback to CPU
  precision: "fp32"  # fp32 for stability, fp16 for speed (if GPU supports)
  gradient_clipping: 1.0  # Prevent exploding gradients
  weight_decay: 0.0001  # L2 regularization
  seed: 42  # For reproducibility
  num_workers: 4  # Data loading parallelism
  pin_memory: true  # Faster GPU transfer

# Quality Gate Thresholds
quality_gates:
  min_sharpe: 1.5
  min_win_rate: 0.55
  max_drawdown: 0.25
  min_trades: 100
  min_profit_factor: 1.3

# Data Configuration
data:
  training_split: 0.7  # 70% training
  validation_split: 0.15  # 15% validation
  test_split: 0.15  # 15% test (holdout)
  sequence_length: 100  # For sequential models (MAMBA, TFT)
  lookback_window: 50  # Technical indicator lookback
  symbols:
    - ES.FUT  # E-mini S&P 500
    - NQ.FUT  # Nasdaq-100 futures
    - ZN.FUT  # 10-Year Treasury Note
    - 6E.FUT  # Euro FX futures

# Checkpoint Configuration
checkpoints:
  compression: "zstd"  # Compress checkpoints to save space
  format: "safetensors"  # Use SafeTensors format (Hugging Face standard)
  max_checkpoints_per_model: 5  # Keep last 5 checkpoints per model
  auto_cleanup: true  # Delete old checkpoints automatically
  validate_checksums: true  # Verify integrity with SHA-256

# Optuna Tuning Configuration (for future tuning runs)
optuna:
  n_trials: 50
  timeout_seconds: 28800  # 8 hours max per tuning session
  study_name_prefix: "foxhunt"
  sampler: "TPE"  # Tree-structured Parzen Estimator
  pruner: "MedianPruner"  # Early stop poor trials
  objective: "sharpe_ratio"  # Optimize for risk-adjusted returns
  direction: "maximize"
  n_jobs: 1  # Sequential trials (GPU memory constraint)
  storage: "postgresql://foxhunt:foxhunt_dev_password@localhost:5432/foxhunt"