The Kelly Criterion Explained: How Much Should You Actually Stake?

Finding a value bet is only half the job. The other half — the part most bettors ignore — is deciding how much to stake. Stake too little and you leave money on the table. Stake too much and a normal losing run wipes out your bankroll before the edge has time to show. The Kelly Criterion is the mathematical answer to that problem, and understanding it is as important as understanding what value betting is.

Why stake sizing matters as much as selection

Imagine you have a coin that lands heads 55% of the time — a genuine edge. If you bet your entire bankroll on every flip, you'll eventually hit a losing streak long enough to go bust, even though the coin is in your favour. Conversely, betting one cent per flip will produce a tiny, glacially slow profit. Neither extreme is rational. The question is: what fraction of your bankroll maximises long-run growth *without* risking ruin?

This isn't a theoretical edge case. Sports bettors with real positive EV blow up their bankrolls all the time because they over-stake. The selection was right; the sizing killed them. Understanding Kelly is the single best safeguard against that outcome.

The Kelly formula in plain language

The formula is: `f = (p · b − q) / b`, where:

• f — the fraction of your bankroll to stake
• p — your estimated probability that the bet wins
• q — the probability it loses (q = 1 − p)
• b — the net odds you receive (decimal odds − 1; so odds of 2.00 → b = 1.00)

Written slightly differently: `f = (p · b − q) / b = p − q/b`. The numerator `p · b − q` is your expected profit per unit. Divide by `b` to normalise it to a stake fraction. When EV is zero, f = 0 — Kelly never tells you to bet on a break-even proposition. The bigger your edge, the larger the recommended stake; the longer the odds, the smaller the recommended stake for the same edge. You can verify any number instantly with our Kelly calculator.

A fully worked example

Notice how the formula scales naturally with your edge. If your model only had the team at 52% (a thinner edge), full Kelly would drop to 4% of bankroll (fractional Kelly: 1%). At 60% it would rise to 20% full Kelly (fractional: 5%). The math adjusts continuously — you don't need different rules for different situations.

Why serious bettors use Fractional Kelly

Full Kelly maximises the geometric growth rate — it's mathematically optimal if your probability estimate is perfectly accurate. The problem: it never is. A model that rates a team at 55% could realistically be off by 2–3 percentage points in either direction. Applying full Kelly to an estimate with that uncertainty means you are systematically over-betting during the periods when your model is wrong.

Fractional Kelly simply multiplies the full Kelly output by a constant less than 1. At 25% (the fraction we use at TheSharpBook), you accept a modest reduction in theoretical long-run growth in exchange for a dramatic reduction in drawdown depth and volatility. Half-Kelly (50%) is another common choice. Research comparing strategies over long bet sequences consistently shows fractional Kelly produces better risk-adjusted outcomes than full Kelly for any bettor whose model carries estimation uncertainty — which is all of us.

There is also a practical ceiling: we cap stakes at 5% of bankroll per bet, regardless of what Kelly says. Very high-EV outliers (above ~18%) are almost always model noise rather than genuine edge — our engine filters those out before they ever reach the staking calculation. Combined, these guardrails mean the system physically cannot over-bet even if a single probability estimate is wildly off.

The critical caveat: garbage in, garbage out

Kelly assumes your probability estimate is correct. If p is wrong, the formula will tell you to over-bet (or, if you underestimate, under-bet). This is not a flaw in Kelly — it's a feature: Kelly is ruthlessly honest about the fact that the quality of your model is the foundation everything else stands on. No staking strategy can rescue a bad probability estimate.

This is the core reason we invest heavily in model calibration: how a betting model works. A well-calibrated model means when we say 55%, we win roughly 55% of those bets over a large sample. Poorly calibrated probabilities — even if they point in the right direction — will cause systematic Kelly over-staking, turning a positive-EV selection process into a losing bankroll.

Before you stake to Kelly on any probability, ask: where does this number come from, and has it been validated on out-of-sample data? Our model pages show the calibration curves and track record. You can also cross-check EV independently using our EV calculator.

Putting it all together

A complete value-betting process looks like this: (1) your model produces a calibrated probability; (2) you compare it against the bookmaker's implied probability; (3) if EV exceeds your minimum threshold (we use 5%), the bet qualifies; (4) Kelly sizes the stake to the actual edge — small edge gets a small stake, large edge gets more; (5) the 25% fractional multiplier and 5% hard cap clip the output to a safe range.

Each piece exists for a reason. Skip the Kelly step and you're guessing at stake sizes. Skip the fractional multiplier and you'll over-bet when your model is slightly off. Skip the hard cap and one noise spike can do real damage. Together they form a system that can survive the inevitable bad weeks and compound steadily over months. Run any scenario through our Kelly calculator to see exactly how each input changes the output.

The same principles apply across every sport we cover — football, baseball, tennis, hockey, basketball. The formula doesn't change; only the probability inputs do. A 5% edge on an NFL spread and a 5% edge on a tennis match winner get sized identically relative to your bankroll, because Kelly doesn't care about sport — only edge and odds.