The Kelly Criterion for Football Predictions — Optimal Bet Sizing for Maximum Growth

June 19, 2026 · 10 min read

You have built a prediction model. You know the probabilities. But how much should you actually stake on each match? The Kelly Criterion, a formula invented at Bell Labs in 1956, answers that question with mathematical precision — and it can transform how you approach football prediction games.

What Is the Kelly Criterion?

In 1956, John Larry Kelly Jr., a researcher at Bell Labs, published a paper describing a formula for sizing bets to maximize the long-term geometric growth rate of wealth. The idea was originally developed for signal noise problems in telecommunications, but gamblers and investors quickly realized its power.

The Kelly Criterion answers a deceptively simple question: given an edge — a situation where your estimated probability of winning differs from what the odds imply — what fraction of your bankroll should you wager to grow it as fast as possible over time?

The formula for a simple win-or-lose bet is:

f* = p − (q / b)

Where f* is the fraction of your bankroll to bet, p is your estimated probability of winning, q = 1 − p is the probability of losing, and b is the decimal odds minus 1 (the net profit on a winning unit bet).

If the formula returns zero, you have no edge — bet nothing. If it returns a negative number, the edge is on the other side. The formula only recommends a positive bet when your estimated probability exceeds the implied probability from the odds.

A Worked Example

Suppose Brazil are playing Switzerland. Your model gives Brazil a 65% chance of winning. The bookmaker offers decimal odds of 1.70 on Brazil (implied probability: 1/1.70 = 58.8%). Your edge is 65% − 58.8% = 6.2 percentage points.

Converting to the Kelly formula: p = 0.65, q = 0.35, b = 0.70 (the net profit on a 1-unit bet at 1.70 odds).

f* = 0.65 − (0.35 / 0.70) = 0.65 − 0.50 = 0.15

Kelly says you should bet 15% of your bankroll on Brazil. That is a substantial allocation — reflecting a genuine edge combined with reasonable odds.

Now consider a different match: Argentina vs Nigeria, where your model gives Argentina a 75% chance but the odds are 1.30 (implied probability 76.9%). Your edge is negative — Kelly returns a negative fraction, meaning you should not bet on Argentina at all, even though they are strong favorites.

Why Kelly Matters for Football Predictions

Most football prediction guides focus on building better models — xG analysis, Poisson distributions, Elo ratings, machine learning. All of that matters. But there is a second question that is just as important: how do you translate a probability estimate into action?

In prediction games like FanPick, every matchday presents a set of decisions. You might have strong opinions on three matches and weak opinions on five others. Kelly gives you a principled way to allocate your confidence — or your virtual points — across those matches based on the size of your edge.

Without Kelly, most people fall into one of two traps:

• Flat betting: Wagering the same amount on every match regardless of confidence. This leaves money on the table when you have a big edge and exposes you to unnecessary risk when your edge is small.
• Gut-feel sizing: Betting more on matches you "feel strongly about" without quantifying why. Emotions distort sizing — people tend to overbet on popular teams and underbet on obscure leagues where edges are larger.

Kelly replaces both approaches with a single formula grounded in information theory. It maximizes the expected logarithm of your bankroll, which is equivalent to maximizing the long-term geometric growth rate. Over hundreds of predictions, the difference between Kelly sizing and flat betting compounds dramatically.

Applying Kelly to Football Match Prediction

Football introduces complications that the basic Kelly formula does not fully address. Matches have three possible outcomes (home win, draw, away win), not two. And prediction games often use point-based scoring rather than binary win/lose payouts. Here is how to adapt.

Step 1: Convert Your Model to Probabilities

Your prediction model — whether it uses Poisson distributions, Elo ratings, xG data, or machine learning — must output probabilities for each match outcome. If your model produces score predictions, convert them to outcome probabilities. For example, if your Poisson model predicts an expected score of 1.8 to 0.9, calculate the probability of each outcome using the Poisson probability mass function.

The quality of your Kelly sizing depends entirely on the quality of your probability estimates. A model that outputs 60% when the true probability is 55% will systematically overbet. This is the single biggest risk of Kelly in practice.

Step 2: Establish Your Odds or Scoring System

If you are sizing actual bets, the bookmaker odds give you b directly. In prediction games without monetary stakes, you need to define what "b" means in terms of points or confidence. One approach:

• Assign each match a "payout" based on how many points you earn for a correct prediction. In FanPick's confidence pool format, you rank your picks by confidence — Kelly helps you decide which matches deserve the highest confidence slots.
• Alternatively, treat Kelly fractions as relative weights. If Match A gets f* = 0.12 and Match B gets f* = 0.04, assign Match A three times the confidence points of Match B.

Step 3: Calculate Kelly for Each Match

For a three-outcome match (home win, draw, away win), use the generalized Kelly formula. For each outcome i:

f*_i = (p_i × b_i − q_i) / b_i

Where p_i is your model's probability for outcome i, b_i is the net payout for that outcome, and q_i = 1 − p_i. Calculate this for each of the three outcomes independently. Only bet on outcomes where f* is positive.

In practice, you will often find that one or two outcomes have positive Kelly fractions while the third is negative or zero. This is your model saying "the odds misprice this outcome."

Step 4: Normalize and Allocate

Once you have Kelly fractions for every match on a matchday, normalize them. If your total Kelly allocation across all matches exceeds your bankroll (or your available confidence points), scale each bet proportionally so the total equals 100%.

This normalization step is critical. Kelly assumes each bet is independent. In football, matches within the same league or matchday can be correlated — a shock result in one match often affects motivation and tactics in subsequent fixtures. Professional bettors typically reduce their total exposure to account for this correlation.

Fractional Kelly: The Practical Choice

Full Kelly — betting exactly the fraction the formula recommends — is mathematically optimal but psychologically brutal. The variance is enormous. A full Kelly bettor can experience drawdowns of 50% or more on their bankroll, even with a genuine edge. These drawdowns are not signs of a broken model; they are an inherent property of maximizing growth rate.

This is why virtually every experienced practitioner uses fractional Kelly:

• Half Kelly (0.5 × f*): Cuts variance by 75% while sacrificing only 25% of the growth rate. The most popular choice among serious bettors and prediction game players.
• Quarter Kelly (0.25 × f*): Reduces drawdowns dramatically. Growth rate is slower, but the emotional toll is much lower. Recommended when your probability estimates have high uncertainty.
• Tenth Kelly (0.1 × f*): Essentially a "confidence-weighted flat bet." Useful when you are not sure your model is well-calibrated but still want to differentiate between strong and weak picks.

The choice of Kelly fraction depends on your risk tolerance and model confidence. If your model has been backtested on thousands of matches and shows consistent calibration, Half Kelly is reasonable. If you are working with a newer model or limited data, Quarter Kelly or below is safer.

Warren Buffett and Bill Gross are both reported to use Kelly-style analysis in their investment decisions, but both are known to use conservative fractional Kelly approaches to manage the inevitable volatility.

Kelly in Prediction Games vs. Real Betting

The Kelly Criterion was designed for repeated bets with real money at stake. Prediction games like FanPick have a different structure — you compete against other players, points are virtual, and the scoring system may reward accuracy differently from simple payout multiplication. Here is how to adapt.

Confidence Pool Strategy

In a confidence pool, you rank your picks from most to least confident. The Kelly fraction naturally provides this ranking. The match with the highest positive f* gets your highest confidence slot. This is superior to gut-feel ranking because it accounts for both the probability of being correct and the "cost" of being wrong (how much you lose relative to what you could gain).

Multi-Outcome Scoring

Some prediction games award different points for exact score predictions, correct outcomes, or correct goal totals. Kelly can handle multi-outcome payouts by treating each possible result as a separate bet with its own probability and payout. The generalized Kelly formula for n outcomes involves solving a system of equations, but a practical shortcut is to calculate Kelly for the most likely outcomes and ignore extreme long-shots.

Tournament vs. League Play

In a long league season, Kelly's long-run optimization property shines — over 380 Premier League matches, for example, the law of large numbers smooths out variance. In a short tournament like the World Cup, with only 7 matches per team, fractional Kelly becomes even more important because you have fewer iterations to recover from drawdowns.

Common Kelly Mistakes in Football Prediction

The Kelly Criterion is only as good as the inputs you feed it. Here are the most common errors:

Overestimating Your Edge

The biggest danger. If your model says a team has a 70% chance but the true probability is 60%, Kelly will recommend a bet size that is far too large. Over many bets, this overconfidence destroys bankroll growth. Always validate your model with calibration plots — if you predict 70% outcomes, roughly 70% of them should actually happen over a large sample.

Ignoring Correlation

Kelly assumes independent bets. Football matches are not independent — injuries, fixture congestion, psychological momentum, and tactical adjustments create correlations. If you are betting on five matches in the same league matchday, your effective risk is higher than Kelly assumes. Reduce your total allocation by 20-30% to compensate.

Betting Every Match

Kelly often recommends zero or near-zero allocation for matches where your edge is negligible. Resist the urge to bet on every game. The most disciplined Kelly practitioners bet on only 15-25% of available matches — those where their model's probability differs meaningfully from the market or consensus.

Using Kelly with Unproven Models

If your prediction model has not been backtested or calibrated, Kelly will amplify its errors. Before applying Kelly sizing, verify that your model has a Brier score below 0.25 (better than random) and shows good calibration across probability bins. A poorly calibrated model with Kelly sizing is worse than flat betting.

Kelly and the Mathematics of Bankroll Growth

The mathematical beauty of Kelly lies in its connection to information theory. Kelly showed that the optimal growth rate of a gambler's bankroll is equal to the mutual information between the gambler's signal (their model) and the actual outcome. In plain terms: the more your model knows that the market does not, the faster you should grow.

Daniel Bernoulli arrived at the same conclusion in 1738 — 218 years before Kelly — through a different route. Bernoulli argued that rational decision-makers should maximize the geometric mean of outcomes, not the arithmetic mean. This resolves the famous St. Petersburg Paradox and leads directly to logarithmic utility, which is mathematically equivalent to the Kelly Criterion.

The practical implication for football predictions: a Kelly bettor with a 3% edge per match will see their bankroll grow at approximately 0.09% per match (the square of the edge, roughly). Over 500 matches, that compounds to a 57% increase. A flat bettor with the same edge grows at roughly half that rate.

Building a Kelly-Enhanced Prediction Workflow

Here is a practical workflow for combining your prediction model with Kelly sizing for each matchday:

1. Generate probabilities — Run your model (Poisson, Elo, xG-based, ML, or ensemble) to get home/draw/away probabilities for every match.
2. Compare to consensus — Check the implied probabilities from bookmaker odds or community predictions. Calculate the edge for each outcome.
3. Apply Kelly formula — Calculate f* for each positive-edge outcome. Discard negative-edge outcomes.
4. Apply fractional Kelly — Multiply all f* values by your chosen fraction (0.25 to 0.5 recommended for most players).
5. Normalize — Scale your bets so the total allocation fits your bankroll or confidence point budget.
6. Execute and track — Make your predictions, record the Kelly fractions, and track results over time. Periodically recalibrate your model.

This workflow turns a raw prediction model into a complete decision system. The model tells you what will happen; Kelly tells you how much it matters.

Key Takeaways

• The Kelly Criterion (f* = p − q/b) calculates the optimal fraction of your bankroll to wager based on your edge — the difference between your estimated probability and the implied probability from odds.
• Fractional Kelly is essential for football. Half Kelly (0.5×) or Quarter Kelly (0.25×) sacrifices some growth rate for dramatically lower variance and drawdowns.
• Kelly ranks your picks by edge, not by certainty. A 55% prediction on an underpriced underdog can produce a higher Kelly fraction than a 90% prediction on a correctly priced favorite.
• Model calibration is everything. Kelly amplifies both your edge and your errors. Validate your model's probability outputs before applying Kelly sizing.
• Only bet when you have an edge. Kelly recommends zero allocation when your model agrees with the market. Discipline — skipping matches where you have no edge — is what separates winning prediction strategies from losing ones.