Position sizing without ruining yourself

Fixed fraction is the only rule that survives a bad run

Fixed-dollar staking has a fatal property: it does not shrink when you are losing, so a drawdown raises the fraction of remaining capital you are risking on every subsequent bet. Fixed-fraction staking shrinks automatically, which is the mechanism, not a side effect. It cannot mathematically reach zero on a series of proportional losses, which is what people mean when they say fractional betting cannot be ruined.

It can still ruin you practically. A 50% drawdown needs a 100% gain to recover, and an 80% drawdown needs 400%. Recovery arithmetic is the constraint most people size against without knowing it, and it is why the useful question is not "what maximises growth" but "what drawdown will I still be able to trade through".

The shape of the curve either side of Kelly

The Kelly fraction on a simple bet is the edge divided by the odds: with a 55% chance at even money, the stake is 10% of bankroll. What matters more than the formula is the shape around it. Growth rises to a peak at the Kelly fraction and falls away on both sides, but not symmetrically. At half Kelly you keep three quarters of the growth rate with half the swing. At double Kelly the growth rate is exactly zero: you have a real edge, you are betting it aggressively, and your expected long-run compound return is nothing at all. Past that, a positive edge loses money.

That asymmetry is the entire argument for fractional Kelly. Undershooting costs you a little growth. Overshooting costs you everything, and the penalty accelerates.

Your edge estimate is the weak link, so size for being wrong

Kelly assumes you know the probability. You do not, you estimated it, and the estimate is the least reliable input in the calculation. Suppose you believe an even-money bet is 55% and it is actually 52%. The true Kelly stake is 4%. Betting the 10% your estimate implied puts you at two and a half times Kelly, past the point where growth turns negative. A three point error, well inside normal estimation noise, converted a winning bet into a slowly losing strategy without changing anything you could observe from the results for a very long time.

The practical fix is to size from the lower bound of your interval rather than the point estimate, then take a fraction of that. Quarter to half Kelly on the pessimistic number is a defensible default and is roughly what surviving practitioners converge on, not because they are timid but because they have all watched what happens on the right-hand side of that curve.

• Estimate the probability, then write down the lower end of what you would defend.
• Compute Kelly on the pessimistic number, not the central one.
• Take a quarter to a half of it.
• Cap any single position independently of the formula.

Simultaneous positions are one position

Kelly is derived for sequential, independent bets. Nothing about it holds when you hold six correlated positions at once. Three trades at 2% each on the same underlying factor are a single 6% bet with the appearance of diversification, and a market that resolves that factor against you settles all three on the same morning.

Size the factor, not the ticket. If you cannot describe what the positions share, assume they share everything, because correlations go to one exactly when it matters. This is also the point where a risk-of-ruin calculation earns its place: it is the only tool in the set that answers the question you actually care about, which is not how fast the account grows but how likely it is to stop existing.