Kelly Criterion bet sizing mathematically maximizes the long-term geometric growth rate of a bankroll, outperforming flat betting by 40-200% in ending bankroll over 1,000 picks with a 55% win rate at -110 odds. The trade-off is volatility: full Kelly produces maximum drawdowns of 30-50% compared to 15-25% for flat betting. The practical solution used by professional bettors and platforms like Olympus Bets is fractional Kelly (25-50% of full Kelly), which captures most of Kelly's growth advantage while keeping drawdowns manageable. The critical prerequisite is accurate probability estimation — Kelly amplifies both good and bad probability estimates, so calibration is non-negotiable.
The Fundamental Question of Bet Sizing
Every sports bettor who has identified a genuine edge faces the same question: how much should I bet? The edge is necessary but not sufficient for profitability. A bettor with a 5% edge who bets 50% of their bankroll on every play will go broke. A bettor with the same 5% edge who bets 0.1% per play will barely grow their bankroll at all. Somewhere between reckless oversizing and overly conservative undersizing lies an optimal strategy.
Two approaches dominate the discourse. Flat betting means wagering the same fixed amount (or percentage) on every pick regardless of edge size. Kelly Criterion means varying bet size proportionally to the perceived edge, betting more when the edge is larger and less when it is smaller. The debate between them is one of the most important in quantitative sports betting, and the math provides a clear answer — with important caveats.
Flat Betting: The Simple Baseline
Flat betting is exactly what it sounds like: every bet is the same size. A bettor with a $10,000 bankroll betting 1 unit ($100) places $100 on every play, regardless of whether the model shows a 2% edge or a 12% edge.
Arguments for Flat Betting
Simplicity. No calculations needed. No probability estimates required. Pick a unit size, bet it, track results. This removes an entire layer of complexity and a major source of potential error (inaccurate probability estimates).
Emotional discipline. Flat betting removes the temptation to chase losses with larger bets or let winners inflate confidence. Every bet is the same mechanical size regardless of recent results or emotional state.
Robustness to bad estimates. Since flat betting does not use probability estimates to determine bet size, it is immune to the damage that miscalibrated probabilities cause. A model that thinks it has a 15% edge when it actually has a 3% edge will not oversize bets under flat betting.
The Mathematics of Flat Betting Growth
With flat betting at a fixed percentage f of bankroll per bet, the expected bankroll after n bets is:
E[Bankroll] = B_0 * (1 + f * EV)^n
Where B_0 is starting bankroll, f is the fraction bet, EV is the expected value per unit bet, and n is the number of bets.
For a 55% win rate at -110 odds (EV = +0.50 cents per dollar bet): E[Bankroll after 1000 bets at 1%] = $10,000 * (1 + 0.01 * 0.05)^1000 = $10,000 * 1.6487 = $16,487
This is linear expected growth in log-bankroll terms, and it works. Over time, a flat bettor with a genuine edge will grow their bankroll. The question is whether they are growing it as fast as possible.
Kelly Criterion: The Mathematically Optimal Approach
The Kelly Criterion was derived by John L. Kelly Jr. at Bell Labs in 1956 in a paper titled "A New Interpretation of Information Rate." Kelly proved that to maximize the expected geometric growth rate of wealth (equivalently, the expected log of wealth), you should bet a fraction of your bankroll equal to:
f* = (b * p - q) / b
Where f* is the optimal fraction of bankroll to bet, b is the decimal odds minus 1 (net profit per unit), p is the true probability of winning, and q = 1 - p is the probability of losing.
For a bet at -110 (decimal 1.909, so b = 0.909), with p = 0.55:
f* = (0.909 * 0.55 - 0.45) / 0.909 = (0.500 - 0.45) / 0.909 = 0.055
Kelly recommends betting 5.5% of bankroll on this bet.
For a bet at -110 with only p = 0.53 (a smaller edge):
f* = (0.909 * 0.53 - 0.47) / 0.909 = (0.482 - 0.47) / 0.909 = 0.013, or 1.3% of bankroll.
The formula naturally prescribes larger bets for larger edges and smaller bets for smaller edges. A 55% edge at -110 gets a 5.5% bet. A 53% edge at -110 gets a 1.3% bet. A 51% edge gets a tiny 0.2% bet. And a bet with no edge (50% at -110) gets f* = -0.05, which means "do not bet" — the formula automatically warns you away from negative-EV situations.
Why Kelly Maximizes Growth
The mathematical proof that Kelly maximizes the geometric growth rate proceeds from the observation that after n bets, your bankroll is:
B_n = B_0 * (1 + f * b)^W * (1 - f)^L
Where W is the number of wins and L = n - W is the number of losses.
Taking the log: log(B_n) = log(B_0) + W * log(1 + f * b) + L * log(1 - f)
The expected growth rate per bet is: G(f) = p * log(1 + f * b) + q * log(1 - f)
Maximizing G(f) by taking dG/df = 0 yields: f* = (b * p - q) / b — the Kelly formula.
This proof, due to Kelly and elaborated by Edward O. Thorp in his landmark work on blackjack card counting, establishes that Kelly is the unique strategy that maximizes the geometric (compound) growth rate. Any other strategy — including flat betting — produces a lower growth rate in the long run. The result is not approximate; it is an exact mathematical optimum.
1000-Pick Simulation: Kelly vs Flat Betting
Mathematical proofs are compelling, but simulations make the difference tangible. We simulated 1,000 sequential bets under controlled conditions to compare the two strategies.
Simulation Parameters
| Parameter | Value |
|---|---|
| Starting bankroll | $10,000 |
| Number of bets | 1,000 |
| True win rate | 55% (varies by scenario) |
| Odds | -110 (decimal 1.909) |
| Flat bet size | 1% of initial bankroll ($100) |
| Kelly bet size | Dynamically calculated per bet |
| Simulations per scenario | 10,000 independent trials |
Scenario 1: 55% Win Rate at -110 (Strong Edge)
| Metric | Flat 1% | Full Kelly | Half Kelly | Quarter Kelly |
|---|---|---|---|---|
| Median ending bankroll | $14,800 | $28,400 | $20,100 | $16,900 |
| Mean ending bankroll | $15,000 | $34,200 | $22,500 | $17,600 |
| Max drawdown (median) | 18% | 42% | 24% | 14% |
| Probability of loss at N=1000 | 6.2% | 8.1% | 5.9% | 5.5% |
| Probability of doubling | 12% | 58% | 38% | 21% |
| Worst case (5th percentile) | $8,200 | $6,100 | $8,500 | $9,100 |
At a 55% win rate, full Kelly produces a median ending bankroll 92% higher than flat betting ($28,400 vs $14,800). The mean is even more dramatic because Kelly's right tail is extremely long — the best Kelly outcomes are far better than the best flat outcomes. But the trade-off is clear in the max drawdown column: 42% vs 18%. And the worst-case scenario (5th percentile) for full Kelly ($6,100) is significantly worse than flat betting ($8,200).
Half Kelly captures most of the upside (median $20,100, or 36% better than flat) while keeping drawdowns comparable (24% vs 18%) and actually having a better worst case ($8,500 vs $8,200). This is why half Kelly is the most common professional recommendation.
Scenario 2: 53% Win Rate at -110 (Moderate Edge)
| Metric | Flat 1% | Full Kelly | Half Kelly | Quarter Kelly |
|---|---|---|---|---|
| Median ending bankroll | $11,600 | $13,800 | $12,500 | $11,900 |
| Mean ending bankroll | $11,800 | $14,900 | $13,000 | $12,200 |
| Max drawdown (median) | 21% | 35% | 22% | 16% |
| Probability of loss at N=1000 | 23% | 26% | 22% | 21% |
| Probability of doubling | 2% | 11% | 5% | 3% |
With a moderate edge, the differences are smaller but the pattern holds. Full Kelly still beats flat betting on growth but with higher drawdowns. Half Kelly remains the sweet spot.
Scenario 3: Overconfident Estimates (Thinks 57%, Actually 52%)
This scenario tests what happens when probability estimates are wrong, which is the most important scenario for real-world betting:
| Metric | Flat 1% | Full Kelly (mis-sized) | Half Kelly (mis-sized) |
|---|---|---|---|
| Median ending bankroll | $10,800 | $8,400 | $10,200 |
| Max drawdown (median) | 22% | 52% | 31% |
| Probability of loss at N=1000 | 38% | 58% | 44% |
This is the critical result. When probability estimates are overconfident (the model thinks 57% but reality is 52%), full Kelly is catastrophically worse than flat betting. The median ending bankroll drops below the starting point ($8,400 vs $10,000), and 58% of trials end in a net loss. Full Kelly amplifies overconfidence into oversizing, which accelerates losses.
Flat betting, immune to probability estimation errors by design, still ends with a modest $10,800 median. Half Kelly splits the difference at $10,200, but even that shows the danger of overconfident estimates plus Kelly sizing.
Why Kelly Wins: The Geometric Growth Argument
The fundamental reason Kelly outperforms flat betting is the difference between arithmetic and geometric growth. Flat betting produces arithmetic growth: each bet adds or subtracts a fixed amount, so the bankroll grows linearly over time. Kelly produces geometric growth: each bet adds or subtracts a percentage, so the bankroll compounds.
Consider two simplified scenarios. A flat bettor starting with $10,000 who wins $100 per bet on average will have $10,000 + (1000 * $100) = $110,000 after 1000 bets. But this is not how betting actually works, because the flat bettor is not reinvesting profits.
A percentage-of-bankroll bettor, by contrast, is always reinvesting. After a win, the next bet is larger (because the bankroll grew). After a loss, the next bet is smaller (because the bankroll shrank). This compounding effect is the engine of geometric growth, and Kelly optimizes the compounding rate.
The mathematical proof shows that in the long run (as n approaches infinity), the Kelly strategy almost surely achieves a higher bankroll than any other strategy. This is not a statement about expected value — many strategies have the same expected value as Kelly. It is a statement about the most likely outcome. The median Kelly bettor outperforms the median flat bettor, and the difference grows over time.
Drawdown Analysis: The Price of Growth
Kelly's superior growth comes at a cost: volatility. Because Kelly bets larger amounts when the perceived edge is large, a string of losses on high-confidence picks produces deep drawdowns.
In our 1000-pick simulation at 55% win rate, the worst single drawdown for full Kelly averaged 42% of peak bankroll. That means a bettor who starts with $10,000 and runs it up to $20,000 might see their bankroll drop to $11,600 before recovering. Psychologically, this is extremely difficult to endure — most recreational bettors would abandon the strategy during the drawdown, locking in losses at the worst possible time.
For comparison, the expected maximum drawdown for flat betting at 1% is approximately 18% of peak. A $20,000 peak might draw down to $16,400. Still painful, but much easier to sit through.
Drawdown Duration
Max drawdown percentage only tells part of the story. Drawdown duration — how many bets it takes to recover from a peak-to-trough decline — matters just as much. In our simulations:
- Full Kelly average recovery time from max drawdown: 180 bets
- Half Kelly average recovery time: 120 bets
- Flat betting average recovery time: 95 bets
At 3-5 bets per day, a full Kelly recovery from max drawdown takes 36-60 days. That is one to two months of betting just to get back to a previous peak. Flat betting recovery takes 19-32 days. The psychological difference is significant.
Fractional Kelly: The Professional Compromise
Edward Thorp, who used Kelly Criterion to beat blackjack and later ran one of the most successful hedge funds in history, recommended betting half Kelly in practice. His reasoning was simple and pragmatic:
- You are never 100% certain of your probability estimates
- Half Kelly achieves 75% of full Kelly's growth rate
- Half Kelly produces roughly half of full Kelly's variance
- The penalty for overbetting (going above Kelly) is much worse than the penalty for underbetting (going below Kelly)
The last point deserves emphasis. The relationship between bet size and growth rate is asymmetric. Betting 50% of Kelly (underbetting) reduces growth by only 25%. Betting 200% of Kelly (overbetting by the same amount) does not merely reduce growth — it can produce negative growth, meaning you slowly go broke even with a genuine edge. This asymmetry is why professionals almost universally err on the side of underbetting.
| Kelly Fraction | Growth Rate (% of Full Kelly) | Variance (% of Full Kelly) | Practical Use Case |
|---|---|---|---|
| 25% Kelly | ~44% | ~25% | Conservative portfolios, uncertain estimates |
| 33% Kelly | ~56% | ~33% | Cautious professionals |
| 50% Kelly | ~75% | ~50% | Industry standard for professionals |
| 75% Kelly | ~94% | ~75% | Aggressive, only with highly calibrated models |
| 100% Kelly | 100% | 100% | Theoretical optimum; impractical for most |
| 200% Kelly | Negative | 400% | Guaranteed ruin over time |
The Calibration Prerequisite
The single most important insight from the simulation results is this: Kelly sizing is only as good as your probability estimates. If your model says 60% and the true probability is 60%, Kelly is optimal. If your model says 60% and the true probability is 53%, Kelly is destructive.
This is why probability calibration is not optional for Kelly users — it is existential. A well-calibrated model's 60% predictions should win approximately 60% of the time. A poorly calibrated model's 60% predictions might win only 54% of the time. Kelly applied to the poorly calibrated model will systematically overbet, treating a small edge as a large one.
At Olympus Bets, we apply Bayesian probability shrinkage before Kelly sizing: model probabilities are pulled 15% toward 50% (the uninformative prior) before computing Kelly fractions. This means a raw model output of 65% is shrunk to 0.65 * 0.85 + 0.50 * 0.15 = 0.6275, or 62.75%, before Kelly is applied. This conservative adjustment has been validated against thousands of historical predictions to prevent overconfidence-driven oversizing.
The calibration pipeline feeds into a unit system that maps Kelly percentages to discrete unit sizes with league-specific caps:
| Kelly % | Units | Risk Level |
|---|---|---|
| 0-1% | 0.5u | Minimum viable edge |
| 1-3% | 1.0u | Standard |
| 3-6% | 1.5u | Above average |
| 6-10% | 2.0u | Strong edge |
| 10-15% | 2.5u | Very strong |
| 15%+ | 3.0u | Maximum (rare) |
League caps prevent oversizing in leagues with higher variance: CBB is capped at 2.5u, soccer and NHL at 2.0u. These caps reflect the empirical finding that model overconfidence is more pronounced in leagues with smaller sample sizes and more variable outcomes.
Real-World Considerations
Bet Limits and Market Access
Kelly theory assumes you can bet any amount at the quoted odds. In practice, sportsbooks limit sharp bettors, reduce maximum wagers, and may close accounts. A full Kelly recommendation of 5% of a $100,000 bankroll is $5,000 per bet, which many books will not accept on non-major markets. This practical constraint pushes bettors toward smaller fractional Kelly amounts regardless of mathematical optimality.
Simultaneous Bets
Standard Kelly assumes sequential bets. When you have five picks on the same night, the naive approach of applying Kelly to each independently risks overexposure. If all five picks are correlated (e.g., all NBA overs on a high-scoring night), a combined loss wipes out a disproportionate chunk of bankroll. Sophisticated Kelly implementations account for correlation between simultaneous bets, typically by reducing individual bet sizes when multiple bets are active.
The Utility Argument
Kelly maximizes the geometric growth rate of wealth, which is equivalent to maximizing the expected logarithmic utility of wealth. But not every bettor has logarithmic utility. A bettor with $10,000 who needs exactly $15,000 for a specific purchase has a different utility function than a professional growing a six-figure bankroll. For specific financial goals, strategies other than Kelly may be more appropriate. Kelly is optimal only for bettors whose objective is to maximize long-run compound growth of their bankroll.
The Verdict: Fractional Kelly Wins
The evidence from both mathematical theory and simulation is clear:
- Full Kelly maximizes long-run growth but produces unacceptable drawdowns for most bettors and is extremely sensitive to probability estimation errors
- Flat betting is simple, robust, and adequate, but leaves significant growth on the table by ignoring edge-size information
- Fractional Kelly (25-50%) captures most of Kelly's growth advantage while maintaining flat-betting-level drawdowns and providing a natural hedge against probability estimation errors
For bettors with well-calibrated probability models (like Monte Carlo simulations with Bayesian calibration), half Kelly is the industry-standard recommendation. For bettors who are less confident in their probability estimates, quarter Kelly or even eighth Kelly still outperforms flat betting while remaining conservative.
The worst possible approach is full Kelly with overconfident estimates. If you are going to use Kelly and your probability estimates have not been rigorously backtested and calibrated, you are better off with flat betting. Kelly is a powerful accelerator — but it accelerates whatever you give it, including overconfidence.
Further Reading
- Kelly Criterion: The Complete Guide — full formula, worked examples, and unit sizing tables
- Kelly Criterion Calculator — interactive tool to compute optimal bet size
- Monte Carlo Simulation Guide — how probability estimates are generated
- Bayesian Calibration Guide — why calibration is essential before Kelly sizing
- Expected Value in Sports Betting — the foundational metric that drives Kelly sizing
- Closing Line Value — measuring whether your edge estimates are real
- Our Methodology — full technical overview of the Olympus Bets platform