Position Sizing (v7)

Three tools, three distinct jobs, applied in sequence. No ambiguity, one answer.

Architecture

KELLY GATE → BLACK-LITTERMAN → CLUSTER CAPS → FINAL WEIGHTS

Tool 1: Kelly Pre-Filter Gate "Does this trade have positive expected value?" Negative Kelly = EXCLUDED. No exceptions. No BL override.

Tool 2: Black-Litterman Allocation "Given my views, confidence, and everything else I hold, what weight?" THE allocation method. Replaces EV×Conv, risk-parity, correlation-Kelly, portfolio-context.

Tool 3: Cluster Caps (Stress Failsafe) "If correlations spike to 1.0 in a crash, are we still safe?" BL output exceeds cap → trim, redistribute. Never retired. Never overridden.

Input

Two modes:

Single stock: /sizing NVDA

• Runs Kelly gate on NVDA
• If passes: re-runs BL for full portfolio with NVDA included
• Shows delta: how does adding NVDA change every other weight?

Full portfolio: /sizing

• Runs Kelly gate on all positions
• Runs BL on all passing positions
• Applies cluster caps
• Outputs final weights

Process

1. Read context:
   • investing/theories/TICKER.md — posterior probability, EV, conviction
   • investing/models/black-litterman.py — BL model with current views
   • investing/philosophy.md — cluster caps, constraints
   • investing/portfolio.md — current holdings
2. Kelly Gate:
   • Pull posterior P(+) from theory card (Bayesian-calibrated)
   • Pull upside/downside ratio from EV distribution
   • Compute Kelly fraction: f* = (p × b - q) / b
   • If f* ≤ 0: STOP. Position excluded. Report why.
   • If f* > 0: PASS. Log Kelly fraction for reference (not for sizing).
3. Black-Litterman:
   • Run python3 investing/models/black-litterman.py --phase [1|2]
   • Or compute inline if views have changed since last run
   • BL inputs:
     • View vector (Q): EV estimates from gauntlet
     • Confidence (Ω): conviction → omega mapping (1/5=0.50, 2/5=0.20, 3/5=0.08, 4/5=0.03, 5/5=0.01)
     • Covariance (Σ): 2yr daily returns, annualised
     • Market equilibrium: cap-weighted implied returns
     • Risk aversion (δ): from regime detection (default 3.5 = cautious)
     • Constraints: max 20% per position, long-only
   • Output: optimal weight per position
4. Cluster Caps:
   • Check BL output against cluster limits:
     • AI Infrastructure (semis, foundry, networking): max 40% of stock allocation
     • Enterprise SaaS: max 40%
     • Consumer/Fintech: max 40%
     • Single position: max 20% of total capital
   • If exceeded: trim to cap, redistribute proportionally to uncapped positions
   • Log any cap-triggered adjustments
5. Apply phase allocation:
   • Phase 1: 45% ETF + 45% stocks + 10% cash
   • Phase 2: 5% ETF + 87% stocks + 8% cash
   • BL weights are within the stock allocation bucket
6. Present results.

Output Format

## Sizing: [TICKER or PORTFOLIO]

### Kelly Gate
| Ticker | P(+)  | Upside | Downside | Kelly f* | Gate |
|--------|-------|--------|----------|----------|------|
| WDAY   | 63.2% | +52.3% | -25%     | 0.34     | PASS |
| HOOD   | 51.1% | +16.5% | -30%     | 0.02     | PASS (marginal) |

### Black-Litterman Allocation
| Ticker | Equilibrium | Your View | Posterior | BL Weight | £ Amount |
|--------|------------|-----------|-----------|-----------|----------|
| WDAY   | +13.9%     | +52.3%    | +16.0%    | 12.1%     | £X       |

### Cluster Cap Check
| Cluster          | BL Total | Cap  | Status |
|------------------|----------|------|--------|
| AI Infrastructure| 18.0%    | 40%  | OK     |
| SaaS             | 24.4%    | 40%  | OK     |

### Final Portfolio (Phase [1|2])
| Ticker | Weight | £ Amount | Role           |
|--------|--------|----------|----------------|
| VWRP   | 45.0%  | £X       | Stability floor|
| WDAY   | 12.1%  | £X       | BL allocation  |
| ...    | ...    | ...      | ...            |
| CASH   | 10.0%  | £X       | IPO reserve    |

### Risk Metrics
Expected return: +X.X%
Volatility: X.X%
Sharpe: X.XX

### BL vs Previous Allocation
[Show deltas, flag moves > 2pp]

Rules

• Kelly gate is binary. Negative = excluded. No negotiation.
• BL output is the answer. Gabriel does NOT see 7 competing numbers and pick.
• Cluster caps cannot be overridden, even by Gabriel. They exist for the regime BL can't model.
• Show full BL math only if Gabriel asks. Default: show results table.
• All amounts in GBP.
• Never auto-execute trades.
• NEVER give financial advice.

Retired Methods (v6)

The following v6 skill files are deprecated. They remain on disk for reference but are NOT run:

• kelly.md → Kelly stays but ONLY as binary gate, not sizer
• bayesian-kelly.md → posterior feeds BL confidence matrix instead
• correlation-kelly.md → BL covariance handles this
• expected-value.md → BL view vector
• risk-parity.md → BL covariance structure
• max-drawdown.md → stress test + cluster caps
• portfolio-context.md → BL constraints