matchedbettors
Try EVSTREAM Free

Last verified: 2026-05-18.

Kelly criterion is a formula for sizing bets proportional to your edge. It tells you what fraction of your bankroll to stake on a given position, given two inputs: the price you are receiving and your estimate of the true probability of the outcome. The formula has a clear rationale, a simple structure, and a well-documented set of limits that matter in practice for AU EV bettors.

Kelly fits inside the broader EV-betting toolkit. See our EV betting Australia guide for the full context.

The Problem Kelly Solves

When you are placing bets with a positive expected value, you face a sizing problem that is not obvious until you have lived through it once.

Bet too little and your edge compounds slowly. A 3% edge on a $10 stake earns $0.30 in expected value per bet. The same 3% edge on a $100 stake earns $3.00. Underleveraged bettors leave a significant fraction of their edge’s growth potential unused. Over hundreds of bets, the difference between staking 1% and 3% of your bankroll on the same edge compounds into a large gap in terminal bankroll size.

Bet too much and you face ruin risk. If you are staking 30% of your bankroll on every positive-EV bet, a short sequence of expected losses at normal variance rates can deplete your bankroll to a level from which recovery is slow or practically impossible. The Kelly formula identifies the stake size that maximises long-run bankroll growth while avoiding the ruin risk that overbetting creates.

These two failure modes exist because of the interaction between edge magnitude, odds, and variance. A 3% edge at 2.00 odds has a different variance profile than a 3% edge at 8.00 odds, even though the expected value per dollar staked is the same. Kelly accounts for this interaction.

The formula was derived by J.L. Kelly Jr. at Bell Labs in 1956. It was initially applied to information theory problems and only later adapted for wagering contexts. The theoretical result is that Kelly staking maximises the geometric mean growth rate of a bankroll over a long sequence of bets with positive expected value. In practice, as described below, most EV bettors use a fraction of full Kelly.

The Formula in Plain English

The Kelly formula is:

f = (b * p - q) / b

Where:

Let’s work through one complete numerical example.

The setup. A bookmaker offers 3.0 on a horse in an AU metro race. You have assessed the horse’s true probability of winning as 40% based on your form assessment and the odds-drift signal you have observed over the morning session.

Identify the variables.

Calculate the Kelly fraction. f = (b * p - q) / b f = (2.0 * 0.40 - 0.60) / 2.0 f = (0.80 - 0.60) / 2.0 f = 0.20 / 2.0 f = 0.10

Full Kelly says stake 10% of your bankroll on this position.

Apply to a $1000 bankroll. Full Kelly stake = $1000 * 0.10 = $100.

The expected value of this bet is: ($100 * 0.40 * 2.0) - ($100 * 0.60) = $80 - $60 = $20. Your expected return is $20 on a $100 stake. The positive expected value confirms that the position is worth taking at this price.

The Kelly fraction tells you not just that the bet is positive expected value (you already knew that, because 3.0 implies 33.3% probability and your estimate is 40%), but how large the position should be relative to your current bankroll to optimise long-run growth.

What happens if the edge disappears. If your true probability estimate was wrong and the horse is actually a 30% chance, then: f = (2.0 * 0.30 - 0.70) / 2.0 = (0.60 - 0.70) / 2.0 = -0.10 / 2.0 = -0.05

A negative Kelly fraction means do not bet. The position has negative expected value at your probability estimate. This is the Kelly formula functioning as a filter, not just a sizing tool.

Fractional Kelly in Practice

Full Kelly staking is theoretically optimal under specific conditions: your edge estimates are accurate, your edge per bet is stable, and you have an infinite time horizon. In practice, AU EV bettors operate under different conditions. Edge estimates are noisy. Sample sizes are finite. Account restrictions can limit stake sizes at some bookmakers. Under real-world conditions, full Kelly is too aggressive for most practitioners.

The standard practice among systematic EV bettors is to use a fraction of full Kelly. The most common fractions are 1/4 Kelly and 1/2 Kelly.

1/4 Kelly applied to the worked example above: Full Kelly stake: $100 1/4 Kelly stake: $100 * 0.25 = $25

On a $1000 bankroll with a 3.0 price and a 40% estimated probability, fractional Kelly says stake $25 rather than $100.

Why use 1/4 Kelly if full Kelly is theoretically optimal? Because the theoretical optimality assumes your probability estimate is correct. In practice, your estimate of a horse’s true probability is itself uncertain. If you estimate 40% but the true probability is actually 35%, your Kelly fraction is too large. Using a fraction of Kelly builds a buffer against the estimation error that is inherent in any real-world EV-betting practice.

The tradeoff is explicit: fractional Kelly accepts slower bankroll growth in exchange for reduced variance and a lower sensitivity to estimation error. Most practitioners find this tradeoff acceptable. The long-run growth rate at 1/4 Kelly is still positive and still compounds meaningfully. What it eliminates is the ruin risk that comes from acting as if your edge estimates are perfectly calibrated.

There is no single correct Kelly fraction. 1/4 Kelly is a common convention, not a mathematical law. Some practitioners use 1/3 Kelly; others use 1/2 Kelly for higher-confidence positions and 1/4 Kelly for lower-confidence ones. The choice should reflect your honest assessment of how accurate your edge estimation is.

Limits and Caveats

Kelly criterion has three significant limits that are particularly relevant for AU EV bettors.

Edge estimation error compounds. The Kelly formula is highly sensitive to errors in the probability input. An overestimate of your edge by 50% can turn a positive-EV sizing decision into a position that contributes to ruin rather than growth over a long sequence. This is the primary reason fractional Kelly is recommended over full Kelly in real-world application. If you consistently overestimate your edge at specific bookmakers or on specific race types, your Kelly-sized positions will be systematically too large.

The practical implication is that tracking your actual results against your pre-bet probability estimates, across a large enough sample to be statistically meaningful (minimum several hundred bets), is essential if you are using any form of Kelly staking. Without this calibration check, you are staking on the assumption that your estimates are accurate rather than on the evidence that they are.

Account restrictions constrain stake size regardless of Kelly. Many AU bookmakers restrict accounts that show a pattern of winning, particularly at small stake sizes on higher-price runners. The restriction may take the form of a lower maximum bet limit or a requirement to request betting approval for stakes above a threshold. This means that even if Kelly says stake $80, your account’s maximum accepted stake for that market might be $15. Account restrictions are an operating reality of AU EV betting that the Kelly formula does not address.

Tools like EVSTREAM are sometimes relevant here because they help identify the bookmakers and market types where your edge is concentrated, which allows you to prioritise accounts with less restrictive terms for the markets where your edge is largest.

Kelly assumes one bet at a time. The standard Kelly formula is derived for a single-bet scenario where you know your bankroll before each bet. In practice, EV bettors often place multiple bets in the same racing session, across different races and different bookmakers. Placing multiple simultaneous Kelly-sized positions can produce a combined stake that is larger than full Kelly would prescribe for the combined position. Some practitioners handle this by dividing their bankroll into sub-pools per bookmaker or per racing session and applying Kelly within each sub-pool independently.

The formula is a tool, not a decision rule that operates autonomously. It requires honest probability estimates, disciplined record-keeping, and calibration against actual results to function as intended.

Continue Reading