How much should you actually stake?

Almost everything written about betting is about which side to take. Almost nothing is written about how much to put on it, and that second question matters just as much as the first.

Two people can agree on every single pick for a year and finish with completely different results, because one of them staked sensibly and the other one did not. Get the sizing wrong and a genuine edge can still lose money, or survive a drawdown that would have wiped out a smaller stake first.

The simplest answer, and why it is not good enough

The easiest rule is a flat unit: pick a size, usually 1% of your bankroll, and bet it on everything. It is simple, and simple is not nothing. But it treats a market we have beaten the closing price on by 7% exactly the same as one we have barely beaten it on at all. That is not risk management, it is refusing to use information you already have.

What the Kelly criterion actually says

In the expected value post we showed that being right is not the same as being paid, and that price is the thing that decides whether an edge turns into money. The Kelly criterion answers the next question: given a real edge, how much of the bankroll should go on it?

The formula itself is short:

stake fraction = edge / (odds − 1)

Bigger edge, bigger stake. Shorter price for the same edge, bigger stake too (a shorter price means less room to be wrong before it costs you). It sounds like the obvious thing to do, and mathematically it is: staked exactly this way, bankroll grows faster over the long run than any other staking rule, proven, not a rule of thumb.

Full Kelly is also more aggressive than almost anyone can stomach. The formula does not know your edge is an estimate, not a fact, and it size up hard on the picks it is most confident about. We tested it on our own ledger: full Kelly did end up ahead, but only after a stretch that took the account down 91% first. Nobody keeps betting through that, so nobody ever gets to the part where it wins.

The practical fix is fractional Kelly: take a quarter of what the formula says and stake that instead. Slower growth, far shallower drawdowns. We tested fractions from an eighth up to double the full formula on our own results, and a quarter came out on top on every measure that matters (final return, return per unit of drawdown, and the chance of finishing behind where you started). Higher fractions bought a fatter best case at a worse typical one, which is the classic overbetting trade, not an improvement.

The part we were still getting wrong

Fractional Kelly needs an edge to work from, and until this week every pick in a given lane got the same number regardless of the market it was in. Every player prop, whatever the sport, was priced as if it carried the same edge. That was never true. Measured against the closing price, a basketball prop kept roughly 70% of its claimed edge. A baseball prop in the same lane kept close to zero. Soccer corner totals ran at 7-10% real edge; a soccer draw, priced in the very same lane, had no edge behind it at all once you controlled for price. One number was doing the job of a dozen.

The fix is to size every market off its own evidence: the same closing-line comparison that proves the whole approach works in the first place, just applied one market at a time instead of once for everything. A market with a short track record gets pulled most of the way back to the overall average until it has earned a number of its own, so a lucky run of five bets in a brand-new market cannot swing its stake around. And every stake still has a hard ceiling, because an estimate is still an estimate.

What that looks like on the real ledger

Same picks. Same prices. Same wins and losses, in the same order they actually happened. The only thing that differs between the two lines is how much was staked on each one, and the gap between them is what a year of routing more of the bankroll toward markets with demonstrated edge, and less toward markets that have not earned it yet, is actually worth.

The honest caveats

This is not a way to manufacture an edge that is not there. It reallocates a fixed amount of edge across markets more sensibly; it cannot exceed what the book has actually earned, and profit at our current sample size is still not proven statistically even though the closing-line edge is. A better staking rule makes a real edge compound faster and a real drawdown shallower. It does nothing for a market that turns out to have no edge at all, which is exactly why the underlying selection work never stops mattering more than the sizing does.

Never stake more than you can afford to lose, whichever rule you use to size it.

Happy betting, Seb