Files
2025-11-30 08:38:26 +08:00

12 KiB
Raw Permalink Blame History

Kelly Criterion Deep Dive

Mathematical foundation for optimal bet sizing under uncertainty.

Table of Contents

  1. Mathematical Derivation
  2. Formula Variations
  3. Fractional Kelly
  4. Extensions
  5. Common Mistakes
  6. Practical Implementation
  7. Historical Examples
  8. Comparison to Other Methods

1. Mathematical Derivation

The Core Question

Problem: What fraction of your bankroll maximizes long-term growth?

Why it matters: Bet too little → Leave money on the table. Bet too much → Risk ruin, high variance.

Logarithmic Utility Framework

Key insight: Maximize expected logarithm of wealth, not expected wealth.

Why log utility?

  • Captures diminishing marginal utility ($1 matters more when you have $100 vs $1M)
  • Makes repeated multiplicative bets additive: log(AB) = log(A) + log(B)
  • Geometric mean emerges naturally (what matters for repeated bets)
  • Prevents betting 100% (avoids ruin)

Derivation for Binary Bet

Setup:

  • Current bankroll: W
  • Bet fraction: f
  • Win probability: p, Loss probability: q = 1 - p
  • Net odds: b (bet $1, win $b net)

Outcomes:

  • Win (probability p): New wealth = W(1 + fb)
  • Lose (probability q): New wealth = W(1 - f)

Expected log utility:

E[log(W_new)] = p × log(1 + fb) + q × log(1 - f) + log(W)

Objective: Maximize g(f) = p × log(1 + fb) + q × log(1 - f)

Finding the Optimum

Take derivative:

dg/df = pb/(1 + fb) - q/(1 - f)

Set equal to zero and solve:

pb/(1 + fb) = q/(1 - f)
pb(1 - f) = q(1 + fb)
pb - pbf = q + qfb
pb - q = f(pb + qb) = fb(p + q) = fb

f* = (pb - q) / b = (bp - q) / b

The Kelly Criterion:

f* = (bp - q) / b

Where:
f* = Optimal fraction to bet
b  = Net odds received
p  = Win probability
q  = 1 - p

Alternative Form

Edge = Expected return per dollar bet = bp - q

Kelly formula: f* = Edge / Odds = (bp - q) / b

Example: p = 60%, b = 1.0 (even money)

  • Edge = 0.6 × 1 - 0.4 = 0.2
  • f* = 0.2 / 1 = 20%

Optimality

Second derivative: d²g/df² < 0 at f = f* → Maximum confirmed

Growth rate: G(f*) maximizes long-run geometric growth

Comparison:

  • f < f*: Lower growth (too conservative)
  • f > f*: Lower growth (too aggressive, variance dominates)
  • f > 2f*: Negative growth (eventual ruin)

2. Formula Variations

Converting Market Odds

Decimal odds (e.g., 2.50): b = Decimal - 1 = 1.50

American odds:

  • Positive (+150): b = 150/100 = 1.50
  • Negative (-150): b = 100/150 = 0.667

Fractional odds (3/1): b = 3.0

Implied probability: Market p = 1/(b + 1)

Multi-Outcome Bet

Horse race: Multiple options, bet on any with positive Kelly

Formula for outcome i:

f_i* = (p_i(b_i + 1) - 1) / b_i

If f_i* > 0: Bet f_i* on outcome i
If f_i* ≤ 0: Don't bet

Continuous Outcomes (Merton's Formula)

Stock market application:

f* = μ / σ²

Where:
μ = Expected return (drift)
σ² = Variance of returns

Example: μ = 8%, σ = 20% → f* = 0.08/0.04 = 2.0 (200%, use leverage)

Reality: Too aggressive, use fractional Kelly → 50-100% more reasonable


3. Fractional Kelly

Why Fractional Kelly?

Problems with full Kelly:

  1. Extreme volatility: Wild swings, can lose 50%+ in bad runs
  2. Model error: If probability estimate wrong, full Kelly overbets dramatically
  3. Practical ruin: 20% chance of 50% drawdown before doubling
  4. Non-ergodic: Most can't bet infinitely many times

Formula

Fractional Kelly = f* × Fraction

Common choices:
- Half Kelly: f*/2
- Quarter Kelly: f*/4 (recommended)
- Third Kelly: f*/3

Growth vs. Variance Trade-off

Strategy Growth Rate Volatility Max Drawdown
Full Kelly 100% 100% -50%
Half Kelly ~75% 50% -25%
Quarter Kelly ~50% 25% -12%

Key: Half Kelly gives 75% of growth with 25% of variance → Better risk-adjusted return

Robustness to Error

Example: You think p = 0.60, true p = 0.55, even money bet

Full Kelly (f = 20%):

  • Growth rate = 0.55×log(1.20) + 0.45×log(0.80) ≈ 0 (breakeven!)

Half Kelly (f = 10%):

  • Growth rate = 0.55×log(1.10) + 0.45×log(0.90) ≈ 0.005 (still positive)

Lesson: Overbetting much worse than underbetting. Fractional Kelly provides buffer.

Situation Fraction Reasoning
Professional gambler 1/4 to 1/3 Reduces career risk
High model uncertainty 1/4 or less Error buffer crucial
High confidence 1/2 to 2/3 Can use more aggression
Institutional 1/4 to 1/3 Drawdown = career risk

Default: Quarter Kelly (1/4) for most real-world situations


4. Extensions

Multiple Simultaneous Bets

Matrix form (N assets):

f* = Σ⁻¹ × μ

Where:
Σ = Covariance matrix
μ = Expected returns vector

Key insight: Correlated bets reduce optimal sizing

Heuristic: Adjusted Kelly = Individual Kelly × (1 - ρ/2), where ρ = correlation

Example: ρ = 0.6, Individual Kelly = 15%

  • Adjusted: 15% × (1 - 0.3) = 10.5%

Correlated Outcomes

Common correlations:

  • Political: Presidential + Senate races
  • Sports: Team championship + Player MVP
  • Markets: Tech stock A + Tech stock B

Extreme cases:

  • ρ = 1 (perfect correlation): Only bet on one
  • ρ = -1 (negative correlation): Bets hedge, can bet more
  • ρ = 0 (independent): No adjustment

Dynamic Kelly

Problem: Probability changes over time (new information)

Process:

  1. Start with p₀, bet f₀*
  2. New information → Update to p₁ (Bayesian)
  3. Recalculate f₁*
  4. Rebalance (adjust bet size)

Consideration: Transaction costs limit rebalancing frequency


5. Common Mistakes

Mistake 1: Full Kelly Overbet

The error: Using full Kelly in practice

Why wrong: Assumes perfect probability estimate (never true)

Impact: Bet 2×f* → Negative growth rate

Fix: Always use fractional Kelly (1/4 to 1/2)

Mistake 2: Ignoring Model Error

The error: Treating probability as certain

Adjustment:

Uncertain Kelly = f* × Confidence

Example: f* = 20%, 80% confident → Bet 16%

Better: Use fractional Kelly (implicitly adjusts)

Mistake 3: Neglecting Bankruptcy

Reality: Finite games + estimation error → real ruin risk

Drawdown stats (full Kelly, p=0.55):

  • 25% chance of -40% before recovery
  • 10% chance of -50% before recovery

Practical bankruptcy: Client fires you, forced liquidation, can't maintain discipline

Fix: Use fractional Kelly, set stop-loss (if down 25%, pause)

Mistake 4: Ignoring Correlation

Example disaster:

  • 10 bets, each Kelly 10%
  • All highly correlated (same theme)
  • Bet 100% total → Single adverse event → Large loss

Fix: Measure correlations, use portfolio Kelly, diversify themes

Mistake 5: Misestimating Odds

Common confusion:

  • Decimal 2.0: b = 1.0 (not 2.0)
  • "3-to-1": b = 3.0 ✓
  • American +200: b = 2.0 (not 200)

Fix: Always convert to NET payout (b = total return - 1)

Mistake 6: Static Bankroll

Problem: Calculate once, never update

Fix: Recalculate before each bet using current bankroll


6. Practical Implementation

Step-by-Step Process

1. Convert odds to decimal:

# Decimal odds: b = decimal - 1
# American +150: b = 150/100 = 1.50
# American -150: b = 100/150 = 0.667
# Fractional 3/1: b = 3.0

2. Determine probability: Use forecasting process (base rates, Bayesian updating, etc.)

3. Calculate edge:

edge = net_odds * probability - (1 - probability)

4. Calculate Kelly:

kelly_fraction = edge / net_odds

5. Apply fractional Kelly:

fraction = 0.25  # Quarter Kelly recommended
adjusted_kelly = kelly_fraction * fraction

6. Calculate bet size:

bet_size = current_bankroll * adjusted_kelly

7. Execute and track:

  • Record: Date, event, probability, odds, edge, Kelly%, bet
  • Set reminder for resolution
  • Note new information

Position Tracking Template

Date: 2024-01-15
Event: Candidate A wins
Your probability: 65%
Market odds: 2.20 (implied 45%)
Net odds (b): 1.20
Edge: 0.43 (43%)
Full Kelly: 35.8%
Fractional (1/4): 8.9%
Bankroll: $10,000
Bet size: $890
Resolution: 2024-11-05

7. Historical Examples

Ed Thorp - Blackjack (1960s)

Application: Card counting edge varies with count → Dynamic Kelly

Implementation:

  • True count +1: Edge ~0.5%, bet ~0.5% of bankroll
  • True count +5: Edge ~2.5%, bet ~2.5% of bankroll

Results: Turned $10k into $100k+, proved Kelly works in practice

Lessons: Used fractional Kelly (~1/2), dynamic sizing, managed "heat" (detection risk)

Princeton-Newport Partners (1970s-1980s)

Strategy: Statistical arbitrage, convertible bonds

Kelly application: 1-3% per position, 50-100 positions (diversification)

Results: 19.1% annual (1969-1988), only 4 down months in 19 years, <5% max drawdown

Lessons: Fractional Kelly + diversification = low volatility, dominant strategy

Renaissance Technologies / Medallion Fund

Strategy: Thousands of small edges, high frequency

Kelly application:

  • Each signal: 0.1-0.5% (tiny fractional Kelly)
  • Portfolio: 10,000+ positions
  • Leverage: 2-4× (portfolio Kelly supports with diversification)

Results: 66% annual (gross) over 30+ years, never down year

Lessons: Kelly optimal for repeated bets with edge. Diversification enables leverage. Discipline crucial.

Warren Buffett (Implicit Kelly)

Concentrated bets: American Express (40%), Coca-Cola (25%), Apple (40%)

Why Kelly-like: High conviction → High p → Large Kelly → Large position

Quote: "Diversification is protection against ignorance."

Lessons: Kelly justifies concentration with edge. Still uses fractional (~40% max, not 100%).


8. Comparison to Other Methods

Fixed Fraction

Method: Always bet same percentage

Pros: Simple, prevents ruin

Cons: Ignores edge, suboptimal growth

When to use: Don't trust probability estimates, want simplicity

Martingale (Double After Loss)

Method: Double bet after each loss

Fatal flaws:

  • Requires infinite bankroll
  • Exponential growth (10 losses → need $10,240)
  • Negative edge → lose faster
  • Betting limits prevent recovery

Conclusion: NEVER use. Mathematically certain to fail.

Fixed Amount

Method: Always bet same dollar amount

Cons: As bankroll changes, fraction changes inappropriately

When to use: Very small recreational betting

Constant Proportion

Method: Fixed percentage, not optimized for edge

Difference from Kelly: Doesn't adjust for edge/odds

Conclusion: Better than fixed dollar, worse than Kelly

Risk Parity

Method: Allocate to equalize risk contribution

Difference from Kelly: Doesn't use expected returns (ignores edge)

When better: Don't have reliable return estimates, defensive portfolio

When Kelly better: Have edge estimates, goal is growth

Summary Comparison

Method Growth Ruin Risk When to Use
Kelly Highest None* Active betting with edge
Fractional Kelly High Very low Real-world (recommended)
Fixed Fraction Medium Low Simple discipline
Fixed Amount Low Medium Recreational only
Martingale Negative Certain NEVER
Risk Parity Low-Med Low Defensive portfolios

*Kelly theoretically no ruin risk, but model error creates practical risk → use fractional Kelly

Final Recommendation: Quarter Kelly (f/4)* for nearly all real-world scenarios.


Return to Main Skill

← Back to Market Mechanics & Betting

Related resources: