bet-sizing

The saved excerpt is a snapshot from when this skill was listed. The external source remains the complete and most current version.

---
name: bet-sizing
description: "Determine how much capital to allocate to individual positions within a portfolio. Use when the user asks about position sizing, the Kelly criterion, fractional Kelly, risk budgeting, or conviction weighting. Also trigger when users mention 'how much to put in one stock', 'maximum position size', 'how concentrated should my portfolio be', 'number of holdings', 'VaR budget per position', 'how big a bet', or ask about scaling position sizes with volatility."
---

# Bet Sizing

## Core Concepts

### Kelly Criterion (Discrete)
For a binary bet with payoff odds b, win probability p, and loss probability q = 1-p:

f* = (b*p - q) / b

where f* is the optimal fraction of wealth to wager. The Kelly criterion maximizes the expected logarithm of wealth (geometric growth rate) over repeated bets.

Properties:
- f* = 0 when edge = 0 (no bet when there is no advantage)
- f* < 0 when negative edge (the formula tells you to bet the other side)
- f* > 0 only when b*p > q (positive expected value)

Note: the reference script's `discrete_kelly` clamps negative Kelly fractions to 0 (no bet) rather than returning a negative value — it does not recommend taking the other side.

### Kelly Criterion (Continuous / Investment)
For a normally distributed investment return with expected excess return mu-r_f and variance sigma^2:

f* = (mu - r_f) / sigma^2

This gives the fraction of total wealth to allocate. For example, an asset with 8% expected excess return and 20% volatility: f* = 0.08 / 0.04 = 2.0 (200% of wealth — implying leverage).

### Fractional Kelly
Full Kelly sizing is theoretically optimal but practically too aggressive because:
- It assumes perfect knowledge of probabilities and payoffs
- It produces large drawdowns (the expected drawdown of full Kelly is significant)
- Estimation error in parameters can turn optimal into catastrophic

Practical approach: use a fraction of Kelly, commonly:
- **Half Kelly (f*/2):** Achieves 75% of the growth rate with substantially lower variance and drawdown risk
- **Third Kelly (f*/3):** Even more conservative; appropriate when parameter uncertainty is high
- **Quarter Kelly (f*/4):** Suitable for highly uncertain estimates

The key insight: the growth rate curve is flat near the peak. Reducing from full Kelly to half Kelly only sacrifices 25% of growth but reduces risk dramatically.

### Risk Budgeting
Allocate risk (not capital) across positions. The total risk budget is the maximum acceptable portfolio risk (e.g., 10% VaR or 5% tracking error).

**VaR-based budgeting:**
- Total VaR budget: e.g., $1M at 95% confidence
- Allocate across positions: Position VaR_i <= allocated VaR_i
- Position VaR = w_i * sigma_i * z_alpha * Portfolio Value

**Tracking error budgeting (for active managers):**
- Total active risk budget: e.g., 4% tracking error
- Allocate across bets: each active bet consumes a portion of tracking error
- Size active positions so that sum of risk contributions equals total risk budget

### Maximum Position Sizes
Hard limits on individual positions to prevent concentration risk:

**Liquidity-based limits:**
- Position < X% of average daily volume (ADV) — common limits: 10-25% of ADV
- Ensures ability to exit within a reasonable time frame (e.g., 5-10 trading days)

**Risk-based limits:**
- Position risk contribution < X% of portfolio volatility (e.g., max 10% of portfolio risk)
- Single position < X% of portfolio value (common: 5% for diversified, 10% for concentrated)

**Regulatory/mandate limits:**
- Mutual fund: no more than 5% in a single name (diversified fund) or 25% (non-diversified)
- Index tracking: weight cannot deviate from benchmark by more than specified amount

### Conviction Weighting
Size positions proportional to the strength of the investment thesis:

- **High conviction (largest positions):** Strong edge, deep research, multiple confirming factors
- **Medium conviction:** Solid thesis but some uncertainty or limited information
- **Low co