Markets coveredMatch OddsCorrect ScoreOver / UnderFirst HalfSecond Half
Statometrics
Module 6 · Lesson 6.2

Fractional Kelly

“Why do the pros bet less than Kelly says?”

Intermediate10 min readBefore this: 6.1 The Kelly criterion

The question

"If Kelly is the stake that grows my bank fastest, why do professionals stake a half or a quarter of it?"

Because full Kelly is only optimal if you know your edge exactly, and nobody does. This lesson shows what fractional Kelly staking costs you in growth, what it buys you in safety, and why the safety is worth far more than it looks.

The idea in one sentence

Staking a fraction of Kelly gives up a little long-run growth in exchange for much smaller swings and protection against overestimating your edge, the mistake almost every bettor makes.

The picture

We simulated 20,000 bettors, each placing 1,000 bets at odds of 3.0 with a genuine 36% win chance (2% commission), starting with a £2,000 bank. The only difference between the groups is how much of Kelly they staked.

Staking Share of bank per bet Median final bank Finished below £2,000 Bank halved at some point Typical worst fall from a peak
Quarter Kelly 0.84% £3,228 9.7% 0.3% 24.5%
Half Kelly 1.67% £4,530 13.6% 8.7% 44.5%
Full Kelly 3.35% £5,928 22.3% 40.2% 72.7%
Double Kelly 6.69% £2,078 48.6% 80.4% 96.1%

Look along the half Kelly row. The typical bettor still more than doubles their bank, their chance of a halving drops from four in ten to under one in ten, and their typical worst fall is 44% rather than 73%.

Now look at double Kelly. Same edge, same bets, and the median bettor finishes almost exactly where they started, after a typical fall of 96%.

Try it · Kelly growth simulator
Kelly fractions to compare
How many bettors
£1,000£10k£100kStartHalf02004006008001,000
━ one bettor each━ median bank
FractionStake per betMedian final bankBelow startHalved at some pointMedian worst fall
¼ Kelly0.84%£3,2279.7%0.2%24.2%
½ Kelly1.67%£4,53014.1%8.3%45.3%
Full Kelly3.35%£5,92820.2%39.2%72.2%
Full Kelly keeps 100% of the growth that staking the true full Kelly would give. The median bettor ends on £5,928, but 39% saw their bank halve along the way.
Each bettor re-sizes to their current bank after every bet. Figures come from 1,000 simulated bettors per fraction (the median bank is also worked out exactly). The chart uses a log scale, so doubling always looks the same height.

Worked Betfair example

The same bet as Lesson 6.1: backing at 3.0 in Over/Under 2.5, 2% commission, bank £2,000. You believe the win chance is 36%. (Illustrative figures; simulation results are from 10,000 to 20,000 simulated bettors.)

  1. Full Kelly on your belief. Net odds 1.96, edge 0.36 × 1.96 − 0.64 = +0.0656, Kelly 0.0656 ÷ 1.96 = 3.35%: a £66.94 first stake.
  2. Half Kelly. 1.67%: a £33.47 first stake. From the table, this keeps about 75% of full Kelly's growth rate.
  3. Now the twist: your estimate is two points too high. The true chance is 34%, not 36%. The true edge is 0.34 × 1.96 − 0.66 = +0.0064, a tenth of what you thought.
  4. True Kelly. 0.0064 ÷ 1.96 ≈ 0.33%. Your "full Kelly" of 3.35% is about ten times the right stake.
  5. What happens to the typical bank over 1,000 bets.
    • "Full Kelly" (really 10 × Kelly): median bank £842, below the start 72% of the time.
    • "Half Kelly" (really 5 × Kelly): median bank £1,695, below the start 60% of the time.
    • "Quarter Kelly" (really 2.5 × Kelly): median bank £1,970, roughly break-even.
  6. The lesson. A small, completely normal error in p turned full Kelly from the fastest growth into steady decline. Fractional Kelly didn't make a thin edge profitable, but it kept the bank alive while you found out the edge was thin.

Verdict: quarter to half Kelly is the professional range for a reason. Your probabilities are estimates, estimates are usually too optimistic, and overbetting punishes that far harder than underbetting.

The formula

Growth at a fraction of Kelly

g(xf∗)g(f∗)≈x (2−x)\frac{g(x f^{*})}{g(f^{*})} \approx x\,(2 - x)
  • f* is the full Kelly share of bank.
  • x is the fraction of Kelly you stake (½ for half Kelly).
  • g is the long-run growth rate from Lesson 6.1.

In plain English: half Kelly keeps about ¾ of the growth, quarter Kelly about 44%, and double Kelly about none. The exact figures in our example were 75.2%, 44.0% and 3.5%.

Swings scale with the fraction

Size of swings∝x\text{Size of swings} \propto x
  • ∝ means "in proportion to".

In plain English: halve your stake and the swings in your bank halve, but the growth only drops by a quarter. That's the best deal in staking.

Chance of your bank ever falling to a share a

P(bank ever falls to a)≈a 2/x−1P(\text{bank ever falls to } a) \approx a^{\,2/x - 1}
  • a is the level you're worried about, such as ½ for "halved".
  • x is your fraction of Kelly.

In plain English: with no time limit, full Kelly halves your bank at some point with probability about ½¹ = 50%. Half Kelly: ½³ = 12.5%. Quarter Kelly: ½⁷ ≈ 0.8%. Our 1,000-bet simulation came in a little lower (40%, 8.7% and 0.3%) because the clock stopped.

The cost of overestimating

ftrue∗=b ptrue−qtruebf_{\text{true}}^{*} = \frac{b\,p_{\text{true}} - q_{\text{true}}}{b}
  • p_true is the real win chance, which you never observe directly.

In plain English: Kelly is very sensitive to p because the edge is a small difference between two big numbers. Two points off p cut the right stake by 90%. Staking a fraction of your estimate is insurance against that error. Shrinking the estimate itself (Lesson 5.2) is the other half of the cure.

Try it

Set true and believed chance both to 36%, odds 3.0, and run full, half and quarter Kelly. Check the medians land near £5,928, £4,530 and £3,228. Then drop the true chance to 34% and run again, and watch full Kelly's lines sink.

Common mistakes

  • Treating full Kelly as the target. It's the ceiling. Above it you lose growth and gain risk; well below it you lose a little growth and gain a lot of safety.
  • Assuming your probabilities are right. Almost every model's edges are overstated on paper. A fraction of Kelly assumes that from the start.
  • Judging a staking plan on the average. Full Kelly's average bank looks wonderful, dragged up by a few huge winners. The median, the typical bettor, is what you'll actually live through.
  • Raising the fraction after a good run. A hot streak is weak evidence of a bigger edge (Lesson 5.2). Change the fraction only on strong, long-run evidence such as sustained closing line value.
  • Forgetting the pain. A 45% fall from a peak is normal at half Kelly. If you couldn't sit through that calmly, stake less. See Lesson 6.4.

Staking is only as good as the edge underneath it: why good models still lose money.

Check yourself

1. Roughly how much of full Kelly's long-run growth does half Kelly keep?
2. You believe your win chance is 36% at 3.0, but it's really 34%. What does staking 'full Kelly' on your belief do?
3. In simulation, what share of full-Kelly bettors (with a correctly estimated edge) saw their bank halve at some point in 1,000 bets?
Key takeaway

Stake a quarter to a half of Kelly. You give up a little growth and buy a lot of safety, above all from your own overconfidence.

Go deeper in the Model Library
Next lesson
6.3 Risk of ruin →
What are the chances I go bust?
18+ only. Educational content, not financial or betting advice. Past results do not guarantee future returns. If gambling stops being fun, get free, confidential help at BeGambleAware.org.
Members