Advanced Kelly: Sizing, Fractional Kelly, and Edge Uncertainty

Why basic Kelly gets you in trouble

Plain Kelly assumes you know your true edge with certainty. You don't. Nobody does. The probability you're betting against is itself an estimate, and that estimate has noise around it. If you bet full Kelly on an estimate that's a few points too high, you're now overbet, and the variance compounds against you faster than the math suggests.

The standard recommendation in serious bankroll management is fractional Kelly, usually one-quarter to one-half. The reason isn't conservatism. It's that fractional Kelly is mathematically optimal once you account for the fact that your edge estimate is itself uncertain. Full Kelly is only optimal if your edge is known. Half or quarter Kelly is closer to right for almost everyone.

This guide covers the standard Kelly formula, the variance math that makes full Kelly painful, the edge-uncertainty correction, multi-bet sizing, and the practical rules most pros actually follow.

The standard Kelly formula

For a single bet at decimal odds o with true win probability p, Kelly says bet a fraction of bankroll equal to:

Worked example. You believe a team has a 55% true win probability. The book is offering -110 (decimal 1.91, so b = 0.91).

So if your bankroll is $10,000, full Kelly says bet $555 on this single wager. That number assumes p = 0.55 is exactly right. If p is actually 0.52 (still a real edge of 1.32%, but smaller), the correct bet was $145, not $555. You overbet by 4x. Lose three or four of these in a row and your bankroll has taken a serious hit you didn't deserve based on your real edge.

The variance problem with full Kelly

Full Kelly maximizes the expected logarithm of bankroll. It does not minimize variance. The growth rate is high, but the path to that growth is brutal. The standard result for a Kelly-sized bettor:

That's the trade-off. Full Kelly is the fastest growth rate in expectation. It also has roughly a one-in-three chance of cutting your bankroll in half before it doubles. Most bettors mentally and emotionally cannot ride out a 50% drawdown without changing their behavior, and changing behavior mid-strategy is how good edges get killed.

Fractional Kelly trades a small amount of expected growth for a dramatic reduction in drawdown probability. Half Kelly captures roughly 75% of the growth rate of full Kelly with a 50% drawdown probability of about 11%. Quarter Kelly captures about 44% of the growth rate with a drawdown probability of about 1.2%. The trade-off is so favorable that almost nobody runs full Kelly in practice.

Edge uncertainty: the real reason for fractional Kelly

The drawdown argument is mostly emotional. The mathematical argument for fractional Kelly is edge uncertainty. If your edge estimate has noise, the optimal Kelly fraction shrinks.

Suppose your edge estimate has standard error σe. The optimal bet size adjusts approximately as:

The intuition: when your edge estimate is noisy relative to its size, you should size as if the edge is smaller than your point estimate. The shrinkage factor is exactly the kind of Bayesian regularization you'd want.

Plug in some realistic numbers. You estimate a 3% edge with a 3% standard error (which is roughly what a serious model produces on out-of-sample data):

Half Kelly. That's where the rule of thumb comes from. If your standard error is comparable to your edge, half Kelly is approximately optimal in the Bayesian sense, not just the conservative sense.

If your edge estimate is half as accurate, with a 6% standard error around a 3% edge:

This is why "use quarter Kelly" is the most common rule of thumb in serious betting circles. For most retail bettors, the noise in their edge estimate is comparable to or larger than the edge itself. Quarter Kelly is the right size.

Multi-bet Kelly with simultaneous wagers

If you place multiple bets at the same time, full Kelly on each bet independently overbets your aggregate exposure. Two uncorrelated full-Kelly bets have an aggregate variance equal to the sum of the individual variances, but full Kelly assumes you have your full bankroll riding on each one.

For uncorrelated simultaneous bets, the correct multi-bet Kelly fraction for each individual bet shrinks. The exact solution requires solving a system of equations; an excellent approximation for n simultaneous independent bets:

Plenty of betting situations have correlated bets even when the events look unrelated. Three NFL games on the same Sunday have a real correlation through weather patterns, betting public sentiment, and shared injury report timing. Two same-team props (over rush yards and over receiving yards for the same RB) have explicit positive correlation.

Practical heuristic: total exposure across all simultaneous open bets should never exceed 25% of bankroll, regardless of what individual Kelly says. If your individual Kelly fractions add up to more than 25%, scale them all down proportionally until they fit.

Worked example: NFL Sunday with three bets

You have $10,000 bankroll. You see three NFL spreads you want to bet, each at -110 (b = 0.91), each with an estimated 3% edge. Standard error on each estimate is also 3%.

If you'd just used full Kelly without adjustment, you'd have bet $330 per game, $990 total, almost 10% of bankroll on a single Sunday's NFL slate. With realistic adjustments for edge uncertainty and mild correlation, the right number is closer to $138 per game.

Common Kelly mistakes

Treating posted vig as part of the edge. Edge in Kelly is the true edge against fair odds, not the difference between your model and the posted line. Always devig the line first, compute true edge against the no-vig probability, and then size off that.

Not scaling Kelly to current bankroll. Kelly says bet f* of your current bankroll, not your starting bankroll. After a loss, your bankroll is smaller, so the dollar amount of the next bet is smaller. After a win, the dollar amount grows. This is what makes Kelly produce geometric growth in the first place. Bettors who use a flat dollar size are not running Kelly, they're running a different strategy.

Adjusting Kelly upward when "due" for a win. No. Each bet is independent. Your current bankroll is what determines your next bet, period.

Compounding Kelly errors. If you're betting full Kelly on overestimated edges, the math works against you exponentially. The bigger your overestimate of edge, the more dangerous full Kelly becomes. This is the strongest argument for fractional Kelly in any context where you're not 100% sure of your edge, which is every context.

Ignoring transaction costs. Books with deposit/withdrawal fees, exchanges with commission, and tax considerations all eat into the edge you can actually capture. The Kelly fraction is computed on net edge after costs, not gross edge.

Practical rules

• Start with quarter Kelly. Move to half Kelly only after you have a long track record showing your edge estimates are accurate.
• Cap total open exposure at 25% of bankroll. If individual Kelly fractions exceed this in aggregate, scale down proportionally.
• Never bet more than 5% of bankroll on any single wager regardless of what Kelly says. The math may suggest 10% on a +EV longshot, but the variance on a single 10% bet is too high to recover from in real time.
• Recompute Kelly fractions against current bankroll before every bet. Don't carry old dollar amounts forward.
• If your edge estimate is based on a model, and your model is new, treat your edge estimate as having a standard error larger than your point estimate. That puts you at quarter Kelly or below.
• Track Kelly performance separately from raw ROI. A bettor with 5% Kelly average sizing and +3% ROI is doing better than a bettor with 1% flat sizing and +5% ROI in absolute dollar terms.

The bottom line on Kelly

Kelly is the right framework for thinking about bet sizing. It's also the framework most often misapplied. Full Kelly is correct only in the limit where you have perfect knowledge of your edge, which is never. Real bettors should run somewhere between quarter and half Kelly, scale to current bankroll continuously, and cap aggregate exposure regardless of what individual bets suggest.

The math behind that recommendation isn't conservative caution. It's the actual answer to the question "what's the optimal bet size when my edge estimate is itself a random variable." The answer is "less than full Kelly," and the more uncertain the estimate, the smaller the right fraction.