Kelly vs Flat Stakes: We Replayed Both on 334 Real Bets
Kelly grows a bankroll fastest, if your probabilities are right. We replayed flat stakes and three Kelly variants on the same 334 bets. Full Kelly turned €1,000 into €431.
Last reviewed November 3, 2026
Key takeaways
- >Kelly fraction = (p × odds − 1) ÷ (odds − 1) of your current bankroll
- >On 334 bets with no real edge: flat €10 ended at €720, full Kelly €431, half €681, quarter €833
- >Kelly multiplies your mistakes: an overestimated edge becomes an oversized stake
- >Most users take a half or a quarter of Kelly, and cap the stake
- >No staking plan creates an edge; it only decides how fast you win or lose
How much should you stake? Most bettors never decide; they bet what feels right. The two systematic answers are flat stakes (the same amount on every bet) and the Kelly criterion (a stake that grows with your edge). We replayed both on the same set of real bets, and the result is a warning about what Kelly needs to work.
Flat stakes
Bet the same amount every time, for example €10, or 1% of your starting bankroll. Simple, easy to track, and your results are easy to judge. Every backtest on this site uses flat stakes for that reason.
The Kelly criterion
Kelly stakes a fraction of your current bankroll that depends on your edge:
Kelly fraction = (p × odds − 1) ÷ (odds − 1)
where p is your estimate of the true probability. Example: odds 2.20, and you believe the chance is 50%: (0.50 × 2.20 − 1) ÷ 1.20 = 8.3% of your bankroll. No edge (p × odds ≤ 1): stake nothing.
In theory, Kelly grows a bankroll faster than any other method if your probability estimates are right. That condition is the whole story.
The replay
We used the 334 "value bets" from What is a value bet?: selections where one venue's price was at least 3% above the margin-free consensus of three others, 14 August – 31 October
- For Kelly, we used that consensus as p. Each method started with €1,000:
Method | Final bankroll | Lowest point | Deepest drop | Biggest stake |
|---|---|---|---|---|
Flat €10 | €720 | €325 | 69% | €10 |
Full Kelly | €431 | €283 | 72% | €35 |
Half Kelly | €681 | €542 | 46% | €20 |
Quarter Kelly | €833 | €740 | 26% | €12 |
These bets had no real edge (they returned −8.4% at flat stakes, inside the luck band). Kelly assumed they did, so full Kelly bet as if each one were a gift, and lost more than half the bankroll. Quarter Kelly lost least, simply because it staked least: on average about a quarter of a percent of the bankroll, against 1% for flat €10.
What this teaches
- Kelly multiplies your mistakes. An overestimated edge becomes an overstated stake. If your p is too optimistic, and almost everyone's is, full Kelly bets too much.
- Most people who use Kelly use a fraction of it: a half or a quarter. It gives up some growth when the edge is real, and loses much less when it is not.
- Swings are enormous either way. Even flat €10 fell to €325 at one point in this replay, over 334 long-priced bets. Any staking plan needs a bankroll that can survive the bad runs.
- No staking plan creates an edge. Without value in the prices, every method loses; the only question is how fast.
A sensible default
- Start with flat stakes of 1% or less of your bankroll, until you have enough results (and a positive closing line value) to believe your edge.
- If you switch to Kelly, use a fraction of it and cap the stake.
- Never raise stakes to chase losses.
How we measured this: the 334 value bets from Edge Terminal's four-venue comparison (last 1X2 price before kickoff, 3% to 30% above the consensus of the other three venues), settled chronologically by kickoff, starting bankroll €1,000; Kelly uses the margin-free consensus as the probability. Snapshot taken 3 November 2026. Nothing here is a tip. Only bet what you can afford to lose.