heurist-finance: Skill for Claude Code

Position Sizing

Shared plumbing and output contract: ../_shared/analyst-core.md. Sizing is the one decision that operates on every trade, and the one public skills reduce to "risk 2%." The truths that matter: sizing errors compound geometrically (oversizing a positive-edge strategy can produce ruin), and every sizing formula's inputs are estimates — so the meta-rule is to size for the uncertainty in the estimate, not the estimate.

The three frameworks (use ≥2, take the smaller answer)

1. Fractional Kelly. Kelly fraction f* = p − q/b (p win prob, q=1−p, b win/loss ratio) maximizes long-run growth if the inputs are exact. They never are — and Kelly is brutally asymmetric to input error: betting 2× true Kelly produces zero long-run growth with wild swings. Practitioner consensus (Thorp; MacLean/Ziemba): ¼ to ½ Kelly, which keeps ~75%+ of the growth at ~25% of the variance. Workflow: estimate p and b from the scenario table (judgmental — label it), compute f*, quote ¼–½ of it, and show the sensitivity ("if your 60% win estimate is really 50%, Kelly goes negative — that's how little room this edge has"). A 3-scenario table collapses to binary Kelly inputs by folding: p = combined probability of the profitable scenarios, b = probability-weighted average gain of those ÷ probability-weighted average loss of the rest — label the collapse.

2. Volatility targeting. Size ∝ target_vol / realized_vol of the position (compute 63-day realized from prices SYM --out FILE; cap the multiplier at 1). Portfolio target for an active retail book: ~10–15% annualized. This is the fix that roughly doubled momentum's Sharpe in the academic record (Barroso & Santa-Clara; Moreira & Muir generalize it) — its real function is forcing smaller size in exactly the high-vol conditions where conviction feels highest. A 60-vol name gets a third of a 20-vol name's dollars at equal conviction, no exceptions by vibes.

3. Stop-based risk budgeting (the field-practical one): position = (account × risk-per-trade) ÷ (entry − invalidation). Risk 0.5–1.5% of account per idea for an active book — but the number that actually governs is portfolio heat: the sum of open risk across positions, ≤ ~5–6%. Heat is what the "2% rule" misses: six uncorrelated 1% risks are 6% heat; six correlated 1% risks are one 6% bet (see below). The invalidation price comes from the thesis (technical-evidence for placement discipline), never from the desired position size — sizing to a stop you moved to afford the size is the tail wagging the dog.