Kelly Criterion: How We Size Bets, and Why Full Kelly Is Dangerous

Most bettors think about bet sizing as a gut call: bet more when you feel good about a pick, less when you don’t. The Kelly Criterion replaces the gut call with a formula. It’s not exotic — it’s been standard in professional gambling and investing since the 1950s — but almost nobody who uses the phrase “bet with Kelly” is actually using full Kelly, and for good reason. Full Kelly is mathematically correct and still capable of wrecking a bankroll that never had a bad decision in it.

The formula

For a bet at American odds, Kelly’s fraction of your bankroll is:

f = (bp - q) / b

Where:

• f = fraction of your bankroll to wager
• b = the net odds you’re getting, expressed as a decimal (win $100 on a $110 bet at -110 means b = 100/110 ≈ 0.909)
• p = your true win probability
• q = 1 − p, your true loss probability

The formula is doing something intuitive once you unpack it: bp is your expected return per dollar risked if you win, q is what you expect to lose per dollar risked, and dividing by b scales the result back into a fraction of bankroll rather than a fraction of the bet. If bp − q is negative — you don’t have a real edge — Kelly tells you to bet nothing. That’s the formula working correctly, not a bug.

Worked example: a 55% edge at -110

Suppose you’ve genuinely identified a bet you’ll win 55% of the time, priced at standard -110 odds (b ≈ 0.9091).

f = (0.9091 × 0.55 − 0.45) / 0.9091
  = (0.500 − 0.45) / 0.9091
  = 0.05 / 0.9091
  ≈ 0.055

Full Kelly says bet 5.5% of your bankroll on that single game. On a $2,000 roll, that’s $110 on one college basketball game. If that sounds aggressive for a “genuine” 55% pick, your instinct is correct — and it’s exactly why almost no serious bettor or fund runs full Kelly.

Why full Kelly’s variance is brutal

Kelly sizing maximizes the long-run growth rate of your bankroll — that’s the actual mathematical guarantee. It says nothing about the ride getting there. We simulated 5,000 seasons of 500 bets each, all at a genuine 55% win rate and -110 odds, sizing every bet at the full Kelly fraction (5.5% of current bankroll):

Read that table twice. A bettor with a real, sustainable 55% edge — better than almost anyone actually achieves against the market — has an 88% chance of watching their bankroll fall by half or more at some point using full Kelly, even though the long-run expectation is strongly positive. Half Kelly gets you most of the growth (1.68x versus 1.99x median) for roughly a third of the pain. That trade is not close.

This is the part that trips people up: a genuine winning edge and a losing streak are not contradictory. Variance doesn’t care that your process is correct. See our bankroll management guide for how often losing streaks show up even for a real 54% bettor across a normal season — the number is higher than most people guess.

Fractional Kelly is the professional standard

Because full Kelly’s drawdowns are survivable in theory (you never bet more than you have) but brutal in practice (very few people keep betting the formula through an 84% drawdown without doubting their edge, changing strategy, or running out of stomach), professional bettors and quant funds almost universally bet a fraction of full Kelly — usually half Kelly (50%) or quarter Kelly (25%). The tradeoff is simple and worth memorizing:

• Half Kelly gets roughly 75% of full Kelly’s expected growth rate for about half the variance.
• Quarter Kelly trades further growth for a much smoother equity curve — useful if your edge estimate is uncertain, which it always is.

There’s a second, quieter reason fractional Kelly wins in practice: your edge estimate is a guess with error bars, not a known constant. Betting a fraction of what the formula says gives you a margin of safety against being wrong about your own edge — which brings us to the mistake that actually blows up bankrolls.

SharpBetz’s approach: units, not raw Kelly fractions

We don’t ask anyone to compute a Kelly fraction on the fly. Our model converts its internal Kelly-fraction estimate for each pick into a simple 1-to-5 unit scale, and it refuses to publish a pick at all below its minimum conviction threshold — a genuine “no edge” result gets discarded, not padded out to hit a daily quota. The unit scale is effectively fractional Kelly with the arithmetic done for you: higher-conviction picks get more units, low-conviction picks get fewer, and the scale caps out well below what full Kelly would ask for on our highest-confidence plays. See our results page for how the unit-weighted record has actually performed against the closing line.

The mistake that makes Kelly dangerous: overestimating your edge

Kelly’s formula is only as good as the p you feed it, and this is where theory meets human nature. Bettors are reliably overconfident about their own edge — everyone’s 55% pick “feels like” 60%, and everyone’s actual long-run win rate is closer to 50-52% than they’d guess before tracking it seriously (see closing line value for a faster way to find out where you actually stand).

Here’s why that overconfidence is fatal rather than just suboptimal. We ran the numbers for a bettor with no real edge at all (true 50% win rate) who mistakenly believes they have an edge and sizes bets with Kelly accordingly:

A small overestimate barely matters. Go from believing you have a 52% edge to believing you have a 55% edge, and your simulated bankroll loss after 100 bets nearly triples. Go to 60%, and 85% of the bankroll is gone. The formula doesn’t fail gracefully when you feed it a bad p — it punishes the error harder than the error itself grew, because the bet size and the per-bet risk both scale up together. That compounding is why “just be honest about your edge” is not a platitude in Kelly betting; it’s the entire game.

What this means for your betting

• Never bet full Kelly on your own edge estimate. Half or quarter Kelly is the professional norm, not a timid compromise.
• A real edge still produces brutal-looking losing stretches. Don’t mistake variance for being wrong.
• The single biggest Kelly risk is overconfidence in p, not the formula itself — round your edge estimate down, not up.
• Track your actual win rate honestly over a large enough sample before trusting any specific edge number you plug into the formula.

The worked examples above use standard -110 odds (b ≈ 0.9091) and the public Kelly Criterion formula. Drawdown and bankroll-multiplier figures are from a 5,000-trial Monte Carlo simulation of fixed-fraction Kelly betting at the stated win rates, not from live account data. This is educational content, not financial advice — bet within your means.