Lesson 4 of 8

The Kelly criterion (and why to use a fraction of it)

Step 3 of 6

Worked example

The 60/40 even-money coin, computed by script on 20 Nov 2026. "After 100 flips" assumes exactly 60 wins, the typical (median) outcome.

Bet each flip Typical growth per flip $1 after 100 flips
10% (half Kelly) +1.52% $4.50
20% (full Kelly) +2.03% $7.49
30% +1.49% $4.37
40% (double Kelly) −0.24% $0.78
50% −3.34% $0.03
  • Above Kelly is strictly worse. 30% grows slower than 20% and swings harder.
  • At about double Kelly, growth turns negative. At 40% you shrink, though each flip's EV is still +8% of the bankroll (0.6 × 40% − 0.4 × 40%).
  • Half Kelly keeps about 75% of the growth (1.52 ÷ 2.03) with far smaller swings.

The swings, in a simulation. Simulated, not real: 20,000 runs of 100 flips each, 60% win chance, even money, the same flips for every bet size. Another random seed moves each figure by about a point.

Bet each flip Ended below the starting $1 Fell more than 50% from a high
10% about 7% about 44%
20% about 18% about 99%
40% a little over half 100%

Even at the "best" size, almost every run halved from its high at some point.

What if your odds are wrong?

Say you think the coin wins 60%, but it really wins 55%. The true Kelly fraction is 0.55 − 0.45 = 10%. You bet 20%:

(1.20)^0.55 × (0.80)^0.45 − 1 = −0.01% per flip → about zero growth

A five-point error in your odds turned the "optimal" bet into one that goes nowhere, with huge swings.

Why you can't use this on a stock

In investing you never know the win chance or the payoff. Returns are not even-money flips, rare disasters (lesson 01) matter more than the average, and a few years of history is a tiny sample. Any "Kelly" number for a stock is built on guesses, and guessing too high is the expensive direction.

So use Kelly for two things only:

  1. A ceiling. Betting more than full Kelly is worse in every way, so no sizing rule should ever go above it.
  2. A warning. Over-betting can turn a good idea into a losing strategy. That's why gamblers often bet less than full Kelly, "to reduce the chance of ruin" and allow for errors in their odds (see "fractional Kelly" on Wikipedia), and why lesson 03's risk limits and caps are the tools you actually use.