Started with flat betting. Moved to percentage staking. Spent two years attempting full Kelly. Settled on something between fractional Kelly and flat betting that I've used for the last decade.
The journey through staking methodologies taught me more about the limits of my own edge estimates than about optimal bankroll management.
The Kelly Criterion is theoretically correct. Bet a proportion of your bankroll equal to your edge divided by the decimal odds minus one.
The problem: Kelly requires an accurate edge estimate. My edge estimates are not perfectly accurate. When Kelly is fed an overestimated edge it prescribes overbetting. Overbetting at scale produces drawdowns that functionally eliminate the bankroll before the edge has time to play out.
Full Kelly on a perceived 6% edge that's actually 3% is not double the optimal bet size.
It's potentially catastrophic depending on the variance distribution.
I use quarter Kelly now. Theoretically suboptimal. Practically survivable.
Fractional Kelly. Specifically half Kelly with a maximum single bet cap.
The reasoning is mathematical.
Full Kelly maximizes the logarithm of expected bankroll growth assuming perfect edge estimates.
My edge estimates are not perfect. They're the output of a model that has been validated over fourteen years but still contains uncertainty.
Half Kelly is optimal if my true edge is approximately half my estimated edge.
This seems conservative. It's the appropriate conservatism for someone who knows their model has uncertainty.
The maximum single bet cap: regardless of Kelly output I don't place a bet exceeding 3% of bankroll.
The cap handles the cases where Kelly recommends a large bet because the edge appears extreme. Extreme apparent edge is more likely to reflect model error than genuine edge that large.
The result: a staking plan that's theoretically suboptimal but practically robust to my own estimation errors.
Used flat betting for twelve years. Recently moved to a mild percentage staking approach.
The argument for flat betting that I held for a long time: my edge estimates aren't precise enough to justify variable staking.
If I'm not confident that this selection has twice the edge of that selection, varying my stake based on that distinction is adding noise rather than signal.
Flat betting treats each selection as equal edge. This is wrong but it's wrong in a consistent and manageable way.
Variable staking treats selections as different edge levels. This is right in theory but wrong in execution if your edge differentiation is inaccurate.
I moved to percentage staking specifically to let the bankroll compound during good periods.
Not because I believe Kelly is correct. Because flat betting on a growing bankroll leaves too much on the table when things go well.
At the exchange professional traders used portfolio-based stake sizing.
Not Kelly per selection. Kelly across the total portfolio of open positions.
The key insight: individual bet Kelly stakes don't account for correlation between positions.
If you have five open positions in correlated markets and Kelly each independently you're dramatically overbetting because the correlations mean the true risk is higher than the independent calculation suggests.
Professional position sizing required understanding the correlation structure of your book, not just the edge on individual selections.
Retail bettors almost never do this because they don't have a book in the professional sense.
But anyone who regularly has multiple active bets simultaneously is implicitly taking on correlated risk they haven't calculated.
The staking plan discussion is usually about individual bets. The meaningful staking question is about the aggregate position.
The correlation problem is also why flat betting has a specific practical virtue.
Flat betting doesn't attempt to optimize stake sizes.
It therefore doesn't make correlation errors.
The suboptimal simple plan avoids the errors that the complex optimal plan makes when executed imperfectly.
The theory of Kelly is correct. The practice of Kelly in the hands of a retail bettor with imperfect edge estimates and no correlation modeling is not obviously better than flat betting and may be worse.
Thirty years of staking plan evolution produces a specific conclusion I'll offer directly.
The staking plan matters less than the discipline with which any plan is maintained.
I've used flat betting, percentage staking, and a modified Kelly across different periods.
The periods of best performance have corresponded not with a specific plan but with the periods of most consistent plan execution regardless of the plan type.
The worst periods have corresponded with plan abandonment.
A mediocre plan executed consistently outperforms an optimal plan executed inconsistently.
The specific plan is less important than whether you follow it when following it is hardest.
Hardest is when you're losing. When you're losing the emotional pressure to deviate from the plan is most intense.
The plan that you can maintain during a losing run is the correct plan for you.
For most people this is a simpler plan than Kelly.
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.
