Integrated Risk Manager

Integrated Risk Manager - Unified Position Sizing (v3.3.0)

Overview

Context

Previously, position sizing was done by three separate systems:

1. GARCHRiskManager - volatility-adjusted sizing
2. CapitalManager - 30% safety buffer, 20% position limits
3. AdvancedRiskManager - VaR, concentration, correlation limits

These systems were disconnected - each applied constraints independently, sometimes contradicting each other. Drawdown scaling was applied AFTER all sizing, reducing effectiveness.

Solution: IntegratedRiskManager

Key Components

from alpaca_trading.risk.integrated_risk import IntegratedRiskManager, PortfolioState

# Initialize with existing managers
integrated_risk_mgr = IntegratedRiskManager(
    garch_manager=garch_mgr,
    capital_manager=capital_mgr,
    portfolio_risk_manager=portfolio_risk_mgr,
    max_drawdown=0.15,         # 15% max drawdown
    use_kelly_sizing=True,     # Kelly-inspired sizing
    half_kelly=True,           # Conservative half-Kelly
)

# Build portfolio state
portfolio_state = PortfolioState(
    total_value=100_000,
    cash_available=50_000,
    current_positions={'AAPL': {'qty': 100, 'price': 150.0}},
    current_drawdown_pct=0.05,  # 5% drawdown
    recent_win_rate=0.55,
    avg_win_pct=0.015,
    avg_loss_pct=0.010,
)

# Get unified position size
result = integrated_risk_mgr.calculate_position_size(
    symbol='GOOGL',
    signal_strength=0.75,
    model_size_mult=0.50,  # From 7-action space
    price=175.0,
    returns=returns_series,
    portfolio_state=portfolio_state,
)

print(f"Final value: ${result.final_value:.2f}")
print(f"Vol regime: {result.vol_regime}")
print(f"Drawdown scale: {result.drawdown_scale:.2f}")

Position Sizing Pipeline

1. Kelly-inspired base size (if enabled)
   • Half-Kelly: kelly_fraction * 0.5
   • Bounded: 5% to 25% of available capital
2. GARCH volatility adjustment
   • Dynamic regime thresholds (percentile-based)
   • Multipliers: low_vol=1.2, normal=1.0, high_vol=0.6, extreme=0.3
3. Correlation penalty
   • Reduces size if correlated with existing positions
   • Uses rolling correlation from returns history
4. Quadratic drawdown scaling
   • Applied BEFORE other sizing (earlier = more effective)
   • Formula: 1.0 - (current_dd / max_dd)^2
5. Capital manager constraints
   • 20% max per position, 30% safety buffer

Key Insights

Dynamic Volatility Thresholds (v3.3.0)

Old approach used static thresholds (10%, 25%, 35% annualized volatility). New approach uses percentile-based classification:

• Looks at historical volatility distribution
• Classifies current volatility relative to history
• More adaptive to market conditions

Graduated Profit Inclusion (Capital Manager Fix)

Old approach excluded ALL unrealized profits from conservative account value. Problem: On winning days, available capital DECREASED despite profits.

New approach (v3.3.0):

# Graduated exclusion based on profit size
if profit_ratio < 0.05:
    profit_exclusion = total_unrealized_profit * 0.50  # 50% excluded
elif profit_ratio < 0.10:
    profit_exclusion = total_unrealized_profit * 0.60  # 60% excluded
elif profit_ratio < 0.20:
    profit_exclusion = total_unrealized_profit * 0.70  # 70% excluded
else:
    profit_exclusion = total_unrealized_profit * 0.80  # 80% excluded

Quadratic Drawdown Scaling

Linear scaling (1 - dd/max_dd) reduces size too gradually. Quadratic scaling (1 - (dd/max_dd)^2) is more aggressive near max drawdown:

• At 5% DD (max 15%): scale = 0.89 (mild reduction)
• At 10% DD: scale = 0.56 (moderate reduction)
• At 14% DD: scale = 0.13 (aggressive reduction)

Failed Attempts

References

• arXiv 2506.04358 - Risk-Aware RL Reward for Trading
• arXiv 2508.16598 - Volatility-Adjusted ATR Sizing
• Kelly Criterion best practices for trading