Risk Management

Risk Management (Deterministic Veto Layer)

The agent with veto power. It can scale exposure DOWN or BLOCK a trade; it can never be overridden by the portfolio manager or by an LLM. Implement the hard rules as plain code with a kill switch, so a model hallucination cannot blow up the book.

Mandatory framing

• Risk controls reduce blowups at the cost of upside; that trade-off is intentional.
• Hard caps (max drawdown, position limits, kill switch) live in deterministic code outside the LLM.
• Not personalized advice.

Layered defenses (apply in order)

1. Volatility target — first line. Size each position inversely to its recent realized vol; scale the whole book to a target portfolio vol (e.g., 10% annualized). Pre-empts most blowups; improves Sharpe more than almost any signal tweak. weight_i ∝ target_vol / realized_vol_i, then renormalize and cap.

2. Drawdown-based de-risking. Scale gross exposure down as drawdown deepens — e.g., linearly from full exposure at −5% drawdown to zero risky exposure at −20%. Mechanical and emotionless.

   if dd > -5%:   risk_scale = 1.0
   elif dd > -20%: risk_scale = (dd - (-20%)) / ((-5%) - (-20%))   # 1.0 -> 0.0
   else:           risk_scale = 0.0

3. CPPI floor (hard drawdown cap). Define a floor (e.g., 85% of capital). Invest multiplier × (NAV − floor) in risky assets, rest in T-bills; as NAV approaches the floor, risky exposure → 0. A time-varying multiplier tied to estimated vol improves it. Risk-parity-then-CPPI is a strong tail-control stack.

4. Trend exits over fixed stops for trend strategies (fixed % stops get whipsawed; exit on close below SMA/channel). Use hard catastrophic stops only as a backstop.

5. Position & concentration limits. Caps per name, per sector, per asset class; gross and net exposure caps. Default examples: ≤25% per asset-class sleeve, ≤10% per single ETF, gross ≤100% (no leverage) unless explicitly approved.

6. Conditional-correlation monitoring. Track the rolling correlation matrix and the effective number of bets. Correlations spike toward 1 in crises — monitor stressed correlation, alert and de-risk when diversification collapses.

Position sizing — Kelly, safely

• Full Kelly produces 50-80% drawdowns and assumes you know true probabilities. **Use ¼–½ Kelly as a

CAP** on the vol-targeted size, never as the primary sizer.

• Need 50-100+ trades before win-rate/payoff estimates are stable enough to trust.

The risk report (contract every cycle)

Always net of costs, in-sample and out-of-sample: CAGR, Sharpe, Sortino, Calmar, max drawdown, current drawdown, realized vol vs target, gross/net exposure, largest position, effective # of bets, time-in-drawdown, turnover.

{
  "verdict": "approve | scale | block",
  "risk_scale": 0.6,
  "current_drawdown": -0.08,
  "portfolio_vol": 0.11,
  "breaches": ["GLD weight 12% < cap 25% OK", "gross 0.78 OK"],
  "kill_switch": false
}

Decision flow

1. Receive target weights from portfolio-construction (already scaled by regime-detection). 2. Apply vol target → drawdown de-risk → CPPI → caps. Compute final risk_scale and adjusted weights. 3. If any hard cap or kill-switch fires → block and route to T-bills + human alert. 4. Emit the risk report; hand approved weights to rebalancing / execution.

Pitfalls

Overfitting risk limits to one history; assuming long-run correlations hold in crises; trusting LLM-estimated probabilities. Test across 2008, 2020, 2022. Keep the deterministic layer simple enough to audit.