Fractional Kelly: Bet Less Than You Should

Last lesson ended on a triumphant note: solve for the peak of the growth curve and you get the Kelly fraction $f^* = p - q/b$ — the single bet size that makes your money compound fastest over the long run. Betting more is strictly worse; betting past $2f^*$ ruins you. So the answer is “bet exactly $f^*$,” right?

Almost nobody who actually manages money does. Ask a professional gambler, a card-counter, a hedge-fund quant — anyone who’s had real money on a Kelly bet — and they’ll tell you they bet half of $f^*$, or less. Not because they can’t do the arithmetic. Because they can, and because they’ve felt what full Kelly does to a real bankroll on a real losing streak. This lesson is about the single most important practical adjustment in the entire framework: deliberately betting a fraction of what the theory says. It sounds like timidity. It’s closer to wisdom.

The half-Kelly bargain — most of the growth, half the swings

Here’s the geometry that makes fractional Kelly work. The growth curve $g(f)$ isn’t a sharp spike you have to hit dead-on; near the top it’s a gently rounded hill. Standing right on the summit ($f^*$) versus standing a bit down the slope ($f^*/2$) barely changes your altitude — because near a smooth peak the curve is essentially flat. Formally, close to the top the growth rate behaves like a downward parabola,

$g(f) \approx \mu f - \tfrac{1}{2}\sigma^2 f^2,$

which peaks at $f^* = \mu/\sigma^2$. Growth falls off quadratically as you move away from $f^*$; the volatility of your log-wealth, meanwhile, scales linearly with $f$. Two different speeds — and that mismatch is the entire bargain.

Run the numbers on our canonical 60/40 even-money coin ($p=0.6$, $b=1$, so $f^*=0.20$):

• Full Kelly, $f=0.20$: $g = 0.6\ln 1.2 + 0.4\ln 0.8 \approx 0.0201$ → +2.0%/round (the peak).
• Half-Kelly, $f=0.10$: $g = 0.6\ln 1.1 + 0.4\ln 0.9 \approx 0.0150$ → +1.5%/round — that’s $0.0150/0.0201 \approx$ 75% of peak growth for half the stake.

Cut your bet in half and you keep three-quarters of the compounding. That’s the headline number to tattoo somewhere: half the bet, three-quarters the growth, half the volatility.

The picture makes the asymmetry obvious. Below $f^*$ the curve is nearly flat — under-betting costs you almost nothing. Above $f^*$ it plunges — over-betting costs you everything. So the smart error is to err low, and half-Kelly is erring low on purpose.

You always overestimate your edge — so full Kelly always over-bets

Now the deeper reason, the one that turns fractional Kelly from “nice trade” into “non-negotiable.” The Kelly formula demands two inputs: the true win probability $p$ and the true payoff odds $b$. You never know them. You estimate them — from a model, from history, from gut — and human estimates of one’s own edge run reliably, almost comically, too high. Traders overrate their signals; gamblers overrate their reads; everyone thinks their edge is bigger than it is.

Here’s why that’s not a minor rounding error but a structural trap. Kelly is a maximum — the top of the hump. Plug in an inflated edge and the formula hands you a bet size that sits to the right of the true peak, out on the ruin-facing slope. And that slope, as we saw, is a cliff. Over-estimating your edge doesn’t cost you a little growth; it pushes you toward $2f^*$, where a genuinely positive-edge game grinds you to zero.

A worked scare: suppose the truth is $p=0.6$, $b=1$, so the true Kelly is $f^*=20\%$. But you’re optimistic and believe $p=0.7$. Your believed Kelly is $f = 2(0.7)-1 = 40\%$ — which is exactly $2f^*$ for the real coin. You’d bet 40% of your bankroll every round, convinced you’re maximizing growth, while your true long-run growth is zero and drifting negative. One overconfident guess and you’ve bet yourself onto the ruin line.

This flips the safety logic entirely. If your edge estimate could be off in either direction, the cost of guessing too low is a little forgone growth on the flat side of the hill. The cost of guessing too high is drawdown and ruin on the cliff side. Faced with an asymmetric penalty like that, you protect yourself by shading your bet down — betting a fraction of your computed Kelly, so that even if your true edge is smaller than you think, you’re still safely on the left side of the peak. Half-Kelly means you can over-estimate your edge by a factor of two and still not over-bet the truth.

Even full Kelly hands you brutal drawdowns

Suppose you could know your edge exactly — no estimation error at all. Should you bet full Kelly then? Still probably not, because full Kelly, even played perfectly, is savagely volatile. Maximizing growth and minimizing heartburn are different objectives, and Kelly optimizes only the first.

The clean fact: under full Kelly, the probability that your bankroll ever drops to a fraction $\alpha$ of its starting value is approximately $\alpha$ itself. So there is about a 50% chance you will at some point be down to half your starting bankroll, and about a 10% chance you’ll at some point touch one-tenth of it — not as a permanent loss, just as part of the ride up. Fifty-fifty to watch half your money evaporate is more than most humans, and essentially all institutions with clients, can stomach without panicking and pulling the plug at the worst moment.

Betting a fraction $c$ of full Kelly makes those drawdown odds collapse fast. In the continuous (diffusion) approximation the probability of ever falling to a fraction $\alpha$ is $\alpha^{(2-c)/c}$ — a bigger exponent as $c$ shrinks, so a smaller probability. Halving $c$ from 1 to $\tfrac12$ takes the “ever halve” odds from 50% to about 12%. Here’s the whole trade laid out on the 60/40 coin:

Read across the half-Kelly row: you surrender a quarter of your growth and in return your chance of ever seeing a 50% drawdown falls from a coin-flip to about one-in-eight, and a 90% wipe-down goes from plausible to nearly impossible. Quarter-Kelly buys even more calm at the cost of roughly half the growth. There is no row where full Kelly’s drawdown profile looks like a good deal for a real person with a real stomach and a finite career.

Fractional Kelly is margin of safety

If “deliberately do less than the optimum, to leave room for being wrong” rings a bell, it should: it’s the margin of safety, the same principle a structural engineer uses when she builds a bridge rated for 40 tons and posts the limit at 10. She isn’t confused about the physics. She’s respecting that loads spike, materials fatigue, and her model of the world is an approximation. The gap between what the structure can bear and what she’ll allow onto it is the safety margin.

Fractional Kelly is that exact move in bankroll form. Full Kelly is the load your growth-math says you can bear. The fraction you actually bet is the posted limit — set deliberately below the theoretical maximum to absorb the two things you can’t control: an edge you’ve over-estimated, and a losing streak worse than your average. Same logic in value investing (buy at a discount to estimated worth so a wrong estimate still leaves you whole), same logic in Kelly (bet below computed optimum so a wrong edge still leaves you solvent).

How much to shave — picking your fraction

So how much should you bet — half, quarter, a tenth? There’s no universal number, but there is a clean way to reason about it. Three dials set the fraction:

• How well you know your edge. A sharp, well-tested, stable edge (a proven card-counting count, a backtested-to-death signal) can bear a bigger fraction — say a half. A vague, thin, or shifting edge (a hunch, a small sample, a market that keeps changing) demands a smaller one — a quarter or less. The fraction is a dial for your confidence in the inputs, not just your appetite for risk.
• How much drawdown you (or your clients) can survive. If a 30% drop makes you quit or triggers redemptions, size so a 30% drawdown is unlikely — which, per the table, means well under full Kelly. Your true constraint is often “what drawdown ends me,” not “what maximizes growth.”
• How correlated your bets are. If you’re running many bets at once that tend to lose together, their combined Kelly is much smaller than the sum of the individual ones — a subject the next lesson opens up. Correlated exposure is a reason to shave harder.

When to use it

Use fractional Kelly essentially always — it’s the practical default, and full Kelly is the edge case, not the other way around. Reach for something near full Kelly only in the rare situation where your edge is genuinely, precisely known and you truly do not care about short-run drawdowns (a purely mathematical toy, or a bankroll so segregated that a total loss of this stake changes nothing about your life). Everywhere real money meets an uncertain edge and a human nervous system, bet half or less. The mnemonic: the theory tells you the most you could bet; wisdom tells you to bet less.