Risk & Psychology · intermediate · 8 min

Why Professionals Bet Fractional Kelly

There is a formula that answers "how big should this position be" exactly. Given an edge — a win rate and a payoff ratio — the Kelly criterion returns the single bet size that maximises the long-run growth rate of the account. Bet more than Kelly and you grow slower while taking on more risk; bet less and you grow slower but more smoothly. It is a genuinely optimal answer to a real question.

And almost nobody who understands it bets the full amount. The working default among people who size professionally is half of what the formula says, sometimes a quarter. This lesson is about why the optimal number is the one you deliberately don't use.

What the formula says

For a bet that wins a fraction p of the time and pays b times what it risks when it wins, the Kelly fraction is:

f* = p − (1 − p) / b

Read it as "the edge, divided by the odds". If you win 60% of the time and your winners are twice the size of your losers (p = 0.6, b = 2), then f* = 0.6 − 0.4 / 2 = 0.4 — Kelly says risk 40% of the account on this trade. That is an enormous number, and the fact that it is enormous is the first clue.

Two properties are worth pinning down:

Kelly scales with the edge. A bigger edge — higher win rate, bigger winners — means a bigger optimal bet. No edge (f* ≤ 0) means the optimal bet is nothing: the formula refuses to size a losing game, and Visor clamps it to zero rather than suggesting you bet the other way.
In risk-per-trade terms, Kelly is always below 100%. Because f* = p − (1 − p)/b is always less than p, and p is below 1, the formula never tells you to risk more than your whole account on one trade. Full Kelly is aggressive, but it is not literally suicidal on a single bet. The damage comes from betting it repeatedly.

Why the optimum is too big to use

Kelly is optimal only if two things are true, and in trading neither is.

First, Kelly assumes you know the edge exactly. The formula takes p and b as given, hard facts. In trading you never have them — you have estimates from a backtest, and estimates from a finite sample of trades are noisy. If your true win rate is 55% but your backtest happened to print 60%, plugging 60% into Kelly sizes you for an edge you don't have. And Kelly is savagely sensitive to this: the growth curve as a function of bet size is flat-topped near the optimum but falls off a cliff past it. Overbetting hurts far more than underbetting. Betting the estimate, when the estimate is biased upward — which backtests systematically are — puts you on the wrong side of that cliff.

Second, full Kelly is violently volatile even when the edge is real. A full-Kelly account routinely draws down 50% or more before compounding pulls it back. That is the mathematically optimal path for maximising growth, but no human sits through a run of it without abandoning the strategy at the worst moment — which converts the optimal sizing into a realised loss. The Risk of Ruin lesson shows the mechanism: the same edge sized large has a real probability of hitting a floor it never recovers from, because the account walks one specific path and the path can fall through the floor before the average arrives.

The fix: bet a fraction

Halving the Kelly fraction is the standard response, and the trade-off is asymmetric in your favour. Betting half Kelly gives up only about a quarter of the growth rate but cuts the volatility of the account roughly in half. You lose a little of the theoretical best-case growth and buy a large reduction in the drawdowns that make a strategy un-runnable — and a large margin of safety against having overestimated the edge in the first place. When your inputs are estimates rather than facts, fractional Kelly isn't timidity; it is the correct adjustment for not actually knowing p and b.

That is why Visor's sizing panel shows half Kelly as the working default, with full Kelly beside it as the ceiling and your strategy's fixed per-trade risk as the third reference point — three sizes to compare, not one number to obey.

Why sizing comes after the gates, never before

Every line above assumes the edge is real. If it isn't, none of this helps — a Kelly size computed from a fictional edge is fiction with a decimal point. A backtest that looked good because you tried fifty variants and kept the winner has, most likely, no edge at all: it is the luckiest of fifty noise draws, and its win rate and payoff ratio will not repeat. Size that with Kelly and the formula will confidently hand you a large number built on nothing.

This is why the sizing panel is gated behind the robustness verdict. A backtest that failed its anti-overfit gates gets no sizing figures — only a note that the edge didn't survive testing, so sizing it is gambling. The honest order is fixed and not negotiable: first establish that an edge survives scrutiny (The Overfitting Trap, Why Most Strategies Fail the Gate); then size it, fractionally; then let the sample play out.

The intuition to keep

Kelly is the growth-optimal size — under numbers you don't actually have. The formula is exact; your inputs are estimates, and biased ones.
Overbetting is punished far more than underbetting. The cost of being past the optimum is steep; the cost of being short of it is mild. When unsure, bet smaller.
Fractional Kelly buys survivability cheaply. Half the bet keeps most of the growth and sheds most of the ruin risk.
No sizing rule rescues a fake edge. Prove the edge first; the gate decides whether there is anything to size at all.