Every staking guide explains the Kelly criterion and then stops, right before the part you actually want, which is whether it does anything.
So here it is run against 51,494 settled positions, the same bets under both schemes, with the result that trips most people up first.
The formula
Kelly sets your stake from your edge and the price:
f = p - (1 - p) / b
p is your probability, b is the profit per unit at the offered price,
and f is the fraction of your bankroll to risk. More edge means a bigger
bet. A better price means a bigger bet. No edge means no bet.
We stake a quarter of that, capped at 25% of bankroll. Quarter Kelly is standard practice and it is not timidity: full Kelly is only optimal if your probabilities are exactly right, and ours are not, which is the whole subject of our page on expected value. Fractional Kelly is what you use when you know your model is imperfect.
In practice that meant an average stake of 3.1% of bankroll, a median of 2.3%, and a maximum of 25%.
The comparison that misleads everybody
Same 51,494 positions, staked both ways:
| Flat, one unit per bet | Quarter Kelly | |
|---|---|---|
| Total profit | +2,011 units | +82.4 units |
| Total risked | 51,494 units | 1,596 units |
| Return on capital risked | +3.91% | +5.16% |
Read the first row alone and flat staking wins by a mile: about 24 times the profit. That is the misreading, and it is almost universal.
Flat also risked about 32 times as much capital. Of course it made more units. It had far more money on the table.
The row that means something is the third one. Per unit actually at risk, Kelly returned 5.16% against flat's 3.91%, which is 32% more efficient on the identical set of bets.
Any comparison of staking plans that comes without the amount risked beside it is not a comparison at all.
Where the advantage comes from
This is the part worth understanding, because it is not magic and it is not guaranteed.
Within any group of bets, flat and Kelly hold exactly the same positions. The only thing that can differ is how much capital each bet receives. So Kelly can only beat flat if the bets it favours are genuinely better.
Sort every position by the stake Kelly assigned it, then look at what those bets actually did:
| Stake decile | Share of Kelly capital | Flat return of those bets |
|---|---|---|
| Largest stakes | 31.0% | +5.13% |
| 2 | 18.1% | +12.85% |
| 3 | 13.6% | +3.45% |
| 4 | 10.6% | +1.82% |
| 5 | 8.3% | -0.60% |
| 6 | 6.4% | +6.32% |
| 7 | 4.8% | +6.04% |
| 8 | 3.5% | +0.72% |
| 9 | 2.4% | -1.19% |
| Smallest stakes | 1.2% | +4.52% |
Kelly put 49% of its capital into the top two deciles, where flat staking would have put 20%. Those bets returned +8.99% against a book average of +3.91%. The smaller half of the book, which flat would have funded at 50%, received only 18%.
That concentration is the entire mechanism, and the table above is worth looking at properly before anyone declares it a law. The relationship is not clean. Only 5 of the ten deciles beat the book average, 2 of them lost money outright, and the very smallest stakes outperformed 5 of the nine deciles above them. If a bigger stake reliably meant a better bet, that column would descend in order. It does not.
What is true is narrower and still useful: the top of the ranking worked. The two deciles Kelly funded most heavily were the two that most outperformed, and because they carried roughly half the capital, that was enough to lift the whole book. The middle of the ranking is noise.
Which is also the risk, stated plainly
Kelly is a bet on your own calibration.
If the biggest edges are real, Kelly amplifies them, and that is what happened here. If the biggest edges are instead where your model is most wrong, Kelly puts the most money into exactly the worst positions and amplifies the damage the same way. It is a leverage multiplier pointed at whichever direction your errors run.
Our largest apparent edges sit on long prices, which are precisely the hardest probabilities to estimate. That is why we stake a quarter and cap at 25%, and it is why we would not run full Kelly on this model.
What it does for risk
Worst peak-to-trough decline, measured on each scheme in its own capital:
- Flat: -566 units, or 1.10% of everything it risked.
- Quarter Kelly: -13.7 units, or 0.86%.
Comparable as a share of capital, which is worth saying because Kelly is often sold as though it removes downside. It does not. What it changes is where your money goes, not whether you lose it.
What we would actually tell you
- Compare staking plans on return per unit risked, never on total profit. Total profit mostly tells you who bet more.
- Fractional Kelly, not full. Full Kelly assumes perfect probabilities. Nobody has those.
- Kelly rewards good calibration and punishes bad calibration, harder than flat staking does in both directions.
- Flat staking is not a mistake. It is the honest default when you are unsure how good your probabilities are, and over this sample it still returned +3.91%.
One partial season across 12 events is enough to show the mechanism and not enough to promise the margin holds. Every position behind these numbers is on our public results page.
Every position PropGolf publishes carries its Kelly stake and its settled outcome, wins and losses alike. See the live board.