Last updated: 15 October 2026
A horse you fancy is 6.0 on the exchange. You reckon it’s got a real 25% chance of winning. The price says about 17%.
So how much do you put on it? Your usual tenner? A big chunk of your bank, because the edge looks huge? Something in between that just feels right?
The short answer: stake in proportion to your edge, and keep it small. The Kelly criterion is the formula for working that out. Feed it a confident guess like the one above and it’ll tell you to put 10% of your bank on one horse. Feed it an edge that’s actually been proven over thousands of runners and it comes out nearer 1% to 3.5%.
What is the Kelly criterion?
It’s a staking formula worked out by John Kelly Jr at Bell Labs in 1956. It tells you what share of your bank to put on a bet so your bank grows as fast as it can over a long run of bets, without ever staking the lot.
Here’s the formula:
Kelly stake = (b × p − q) ÷ b
- b is the decimal odds minus one, so what you win on a £1 stake
- p is your chance of the bet winning
- q is your chance of it losing, which is 1 minus p
Put simply, it’s your expected profit on every £1 staked, divided by b.
If the answer comes out at zero or below, Kelly is telling you not to bet. That on its own is worth having, because it’s a check on whether a bet is value at all.
Try it on the horse at the top. b is 5, p is 0.25 and q is 0.75. That’s (5 × 0.25 − 0.75) ÷ 5, which comes to 0.10. Kelly says put 10% of your bank on it.
Hold on to that number.
The catch is the p
Everything rides on p, your chance of the horse winning, and Kelly takes your number as gospel. If your p is too high, Kelly doesn’t argue with you. It just makes the stake bigger.
And look at what that 25% is claiming… the price says 17%, so you’re saying this horse wins half as often again as the market thinks. That’s a massive edge.
So what does a proven one look like?
What a proven edge looks like
Race Advisor’s Race Predictor gives every runner on the card a chance of winning. I took the top fifth of runners by that chance, across over 1.4 million UK runners from 2013 to 2026. That’s over 250,000 horses.
They won 25.94% of the time. Going by their Betfair Starting Prices, about 24 in every 100 should have won… and about 26 did.
That’s an A/E of 1.069. A/E is actual winners divided by the winners the prices expected, so 1.069 means about 7% more winners than the market said. There’s more on it in what an A/E ratio is.
It’s a real edge, measured over a huge number of runners. It’s also history, and an edge that held for thirteen years can still shrink. But notice how small it is. About 7% more winners than the prices expected, not 50%.
That’s normal. The bookies and the exchange are very good at pricing horses, because their whole business depends on it. Beating them by a few per cent, year after year, is a proper edge.
What Kelly says on a real edge
Put that 7% edge into Kelly and the stakes come right down. If every horse in the group carries the same edge, the Kelly stake works out like this:
| Price | Kelly stake | On a £1,000 bank |
|---|---|---|
| 3.0 | about 3.5% | £35 |
| 5.0 | about 1.7% | £17 |
| 8.0 | about 1% | £10 |
That’s before commission. Betfair takes a cut of your winnings, and on an edge this thin the cut eats a good share of it.
Now go back to the horse at the top. Same price, 6.0. On your own 25% guess, Kelly said 10% of your bank, so £100. On the edge this group has actually shown, it’s about 1.4%, so £14. Roughly seven times smaller.
Here’s another one. Two of our ratings, PR and VDW, each rank every runner in a race. When both put the same horse top, it won 36.93% of the time across over 65,000 runners, at prices averaging about 3.4. Again, that’s what happened, not a promise about next season. The A/E there is 1.021, so about 2% more winners than the prices expected. Kelly makes that just under 1% of your bank before commission. At a 5% commission rate the edge is gone altogether, and Kelly says don’t bet.
Why full Kelly still hurts
Even when your p is spot on, full Kelly is a rough ride. On the maths behind it, with p exactly right and a long run of separate bets, there’s about an even chance your bank halves at some point along the way. That’s a lot to sit through, and it’s hard to stick to a plan when it happens.
If your p is too high, it gets worse fast. Stake twice the right Kelly amount and your bank, on average, stops growing. Stake more than twice and it shrinks, even with a real edge.
Say you think you’ve got that 7% edge on an 8.0 shot, so you stake 1%. But the real A/E is 1.03, not 1.069. The right stake was about 0.4%. You’ve been staking more than double, which is exactly where the bank starts going backwards.
And the 25% guess at the top of this post? That had you staking about seven times what the proven edge supports.
How to use Kelly without wrecking your bank
The usual fix is to bet a fraction of the Kelly stake, half or a quarter. You give up some growth, and in return the ride is far smoother and there’s room for being wrong about your edge.
On the 7% edge at 8.0:
| Share of bank | On a £1,000 bank | |
|---|---|---|
| Full Kelly | about 1% | £10 |
| Half Kelly | about 0.5% | £5 |
| Quarter Kelly | about 0.25% | £2.50 |
Small, isn’t it? On a real edge, Kelly stakes look a lot like sensible level stakes. So if Kelly ever tells you to put 10% of your bank on one horse, check your p before anything else. It’s almost certainly too high.
Three things make it work:
- Get p from data, not feel. Use a rating you’ve tested, like the Race Predictor, never a gut-feel percentage. If you can’t say how many bets a number’s based on, it doesn’t go in the formula.
- Bet a fraction. Half Kelly at most. I’d start at a quarter.
- Track every bet. Winners, losers, the losing runs and how far your bank drops, not just the A/E. The bet tracker on Race Advisor does this, and a spreadsheet works too. A good A/E after a few hundred bets is worth watching, but it isn’t proof. On an edge as small as 7%, it takes a few thousand bets before it stands out from luck, and more if you bet at bigger prices. One bad run doesn’t prove the edge has gone, either.
Is it worth it over level stakes?
Not until you know your edge. Start on level stakes, the same amount on every bet, and keep tracking. Until your records run into the thousands and your A/E is still clearly above 1, not just a whisker over it, you’d just be putting bigger money on guesses.
Once they do, move to a small fraction of Kelly. And keep checking it on the bets you’re placing now, at the prices you actually get, because an edge that was there last year can fade.
What this doesn’t cover
Everything here is one bet at a time. Backing two horses in the same race, or having several bets running at once, needs a different sum, and I haven’t gone into that here.
See it on today’s racing
Open today’s race cards and look at the Race Predictor tab on any race. It gives every runner a chance of winning, so you can set each one against what its price says.
On a single horse the two can be miles apart. Don’t size your stake from that gap. Across thirteen years the whole top fifth beat the market by about 7%. That’s the number to put into Kelly.
