release: paperium v2 - LSTM

Author: snowfluke

README.md

@@ -1,5 +1,6 @@
 """
 IHSG Quantitative Trading Model Configuration
+Refactored for LSTM + Triple Barrier Labeling (Paper Implementation)
 """
 from dataclasses import dataclass, field
 from typing import List
@@ -10,109 +11,65 @@
 @dataclass
 class DataConfig:
     """Data fetching configuration"""
-    # Stock universe - Comprehensive Liquid IHSG Universe (IDX80 + Kompas100 + Liquid Growth)
+    # Stock universe
     stock_universe: List[str] = field(default_factory=lambda: IDX_UNIVERSE)
 
     # Data storage
     db_path: str = "data/ihsg_trading.db"
 
     # Historical data settings
-    lookback_days: int = 365 * 5  # 5 years of history
-    min_data_points: int = 252  # Minimum data points for calculation
-
-
-@dataclass
-class SignalConfig:
-    """Signal generation parameters"""
-    # Technical indicators
-    rsi_period: int = 14
-    macd_fast: int = 12
-    macd_slow: int = 26
-    macd_signal: int = 9
-    bb_period: int = 20
-    bb_std: float = 2.0
-    atr_period: int = 14
-    
-    # Mean reversion
-    zscore_period: int = 20
-    zscore_entry_threshold: float = 2.0
+    lookback_days: int = 365 * 5  # 5 years
+    min_data_points: int = 252
 
-    # Momentum
-    momentum_periods: List[int] = field(default_factory=lambda: [5, 10, 20])
+    # Sequence generation
+    window_size: int = 100  # From Paper: Optimal Window Size
 
 
 @dataclass
 class MLConfig:
-    """Machine learning model configuration"""
-    # Training windows (5 years total - using more of your 5-year historical data)
-    training_window: int = 1008  # 4 years of trading days for training
-    validation_window: int = 252  # 1 year of trading days for validation
-    validation_split: float = 0.2  # Deprecated, use validation_window instead
-
-    # Prediction target (aligned with max_holding_days for day trading)
-    target_horizon: int = 5  # Predict 5-day forward return (matches max hold)
-
-    # XGBoost parameters (Conservative Defaults with Strong Regularization)
-    # Note: Based on training results, 50 trees is the sweet spot before overfitting
-    # Regularization allows using more trees without overfitting
-    n_estimators: int = 500
-    max_depth: int = 4
-    learning_rate: float = 0.05
-    min_child_weight: int = 5
-
-    # Regularization parameters (prevent overfitting)
-    subsample: float = 0.7  # Use 70% of data per tree (randomness prevents overfitting)
-    colsample_bytree: float = 0.7  # Use 70% of features per tree
-    gamma: float = 0.2  # Minimum loss reduction required to split (conservative)
-    reg_alpha: float = 1.0       # Stronger L1 regularization (feature selection)
-    reg_lambda: float = 2.0      # Stronger L2 regularization
-
-    use_gpu: bool = False
+    """Machine learning model configuration (LSTM)"""
+    # Model Architecture (Paper: Hidden Size 8 was optimal)
+    input_size: int = 5  # Open, High, Low, Close, Volume
+    hidden_size: int = 8
+    num_layers: int = 2  # Starting with 2 stacked LSTMs
+    dropout: float = 0.0
+    
+    # Training
+    batch_size: int = 64
+    learning_rate: float = 0.001
+    epochs: int = 50
+    patience: int = 10  # Early stopping
+    
+    # Triple Barrier Labeling (Optimized for IHSG)
+    # Found via optimization: Horizon=5, Barrier=3.0%
+    tbl_horizon: int = 5
+    tbl_barrier: float = 0.03
+    num_classes: int = 3  # 0: Loss, 1: Hold, 2: Profit
 
-    # Feature settings
-    feature_lags: List[int] = field(default_factory=lambda: [1, 2, 3, 5, 10, 15, 20])
 
 @dataclass
 class ExitConfig:
     """Exit strategy configuration"""
-    stop_loss_atr_mult: float = 2.0       # Tighter stop (2.0x is standard for swing)
-    take_profit_atr_mult: float = 5.0     # Reduced from 10.0. (4-6x is a "Home Run" in 5 days)
-    trailing_stop_atr_mult: float = 2.5   # Trail loosely to let winners run
-
-    # Time-based exit (hold longer for bigger moves)
-    max_holding_days: int = 5
+    # Legacy ATR-based exits (kept for PositionManager compatibility)
+    # TBL barrier is 3%, so we map roughly to that. 
+    # If ATR is ~2%, then 1.5x ATR = 3%.
+    stop_loss_atr_mult: float = 1.5
+    take_profit_atr_mult: float = 1.5
+    signal_threshold: float = 0.5
 
-    # Fixed stops (fallback for extreme cases)
-    max_loss_pct: float = 0.08    # Tighten max loss to 8% (preservation is key)
-    min_profit_pct: float = 0.15  # Realistic 15% upside target for 1 week
- 
-    signal_threshold: float = 0.55  # Minimum ML signal to enter/hold position
 
 @dataclass
 class PortfolioConfig:
-    """Portfolio management configuration"""
-    # Total portfolio value for sizing calculations
-    total_value: float = 100_000_000  # Default 100M IDR
-    
-    # Position sizing
-    max_positions: int = 8             # Reduced: Focus capital on best ideas
-    base_position_pct: float = 0.125   # 12.5% per trade
-    max_sector_exposure: float = 0.25  # 25% max per sector
-    
-    # Liquidity filter
-    min_avg_volume: int = 2_000_000  # 2M shares minimum
-    min_market_cap: float = 2e12  # 2 Trillion IDR minimum
-    
-    # Risk management
-    max_portfolio_volatility: float = 0.20  # 20% annual vol target
-    max_correlation: float = 0.6  # Avoid highly correlated positions
+    """Portfolio simulation settings"""
+    total_value: float = 100_000_000.0  # Default 100M IDR
+    max_positions: int = 10
+    risk_per_trade: float = 0.02  # 2% Risk per trade (Optional usage)
 
 
 @dataclass
 class Config:
     """Master configuration"""
     data: DataConfig = field(default_factory=DataConfig)
-    signal: SignalConfig = field(default_factory=SignalConfig)
     ml: MLConfig = field(default_factory=MLConfig)
     exit: ExitConfig = field(default_factory=ExitConfig)
     portfolio: PortfolioConfig = field(default_factory=PortfolioConfig)