OLIVER SCOTT
Projects / Trading · Risk

Bet sizing: Kelly and estimation error

probabilityMonte Carlorisk
AbstractHow much should you bet when you have an edge? The Kelly fraction maximises long-run growth, but it assumes you know your edge exactly. I simulate 10,000 bettors over 1,000 bets at fractions of Kelly, check the simulation against closed-form results, and then measure what happens when the edge must be estimated from a limited track record.
10%Kelly fraction for a 55% even-money bet
150×median wealth after 1,000 bets at full Kelly
-0% vs 57%growth kept: full vs half Kelly, edge from 200 bets
11.3%P(halving from start) at half Kelly; theory 12.5%

01Growth versus fraction

g(f) = p·ln(1 + b·f) + (1 − p)·ln(1 − f) f* = p − (1 − p)/b
Growth curve and estimation error
Fig. 1 Left: expected log growth peaks at the Kelly fraction and turns negative past twice Kelly, despite the positive edge. Right: when p is estimated from past bets, betting the full estimated Kelly fraction can destroy growth; fractional Kelly is robust.

02Two kinds of drawdown

FractionfMedian wealthP(end below start)P(ever below ½ start): sim / theoryP(50% fall from a peak)
0.25×0.0258.92×0.3%0.6% / 0.8%9%
0.5×0.05042.63×0.9%11.3% / 12.5%86%
1.0×0.100149.66×6.0%47.4% / 50.0%100%
1.5×0.15041.91×21.4%74.8% / 79.4%100%
2.0×0.2000.87×51.2%90.4% / 100.0%100%

My first check compared the simulated chance of a 50% drawdown with the closed form 0.52/c − 1 and found 86% against 12.5% at half Kelly. Neither was wrong: the closed form is the chance of ever halving from the starting stake, which the simulation reproduces (11.3%). The 86% is the chance of falling 50% from any peak over 1,000 bets, because every new high is a fresh chance. Both are now tested.

03Estimation error

With a true edge of 53% estimated from 200 past bets, full Kelly keeps -0% of the growth that perfect knowledge would give, and half Kelly keeps 57%. With only 50 past bets, full Kelly has negative expected growth. Overestimating the edge costs far more than underestimating it, which is why practitioners size at a fraction of Kelly.

04What I'd do next

  • Multi-asset Kelly with correlated bets and leverage limits.
  • Bayesian sizing that shrinks the estimated edge by its uncertainty.