DawgHousePicks
Every serious sports bettor eventually encounters the Kelly criterion. The formula promises to tell you the optimal fraction of your bankroll to wager when you have an edge. Most quit after their first full-Kelly drawdown. The few who stay with the method use a quarter of what the formula recommends. The adjustment is not timidity. The adjustment reflects one fact about the real world: your estimated edge sits well above your true edge.
This article works through the full math: the formula derivation, the growth-rate curve proving overbetting destroys wealth, the variance-reduction arithmetic behind quarter-Kelly, and the estimation-error argument making quarter-Kelly the default for anyone betting without a perfect model. An interactive bankroll simulator at the end lets you compare all three curves, full, half, and quarter Kelly, on your own numbers.
The formula and the actual objective
John L. Kelly Jr. derived the criterion in 1956 while working at Bell Labs on information transmission over noisy channels. The betting application follows directly. The problem has the same structure. Given a repeated game where you hold a genuine informational advantage (your probability estimate beats the price implied), how much should you bet each round to maximize long-run wealth?
The answer comes from maximizing the expected logarithm of wealth, not the expected dollar amount. This distinction is everything. Maximizing expected dollars leads to catastrophic overbetting. "Bet everything" maximizes expected value on any +EV play, but guarantees ruin in finite time through natural variance. Maximizing expected log-wealth produces a growth-optimal fraction balancing upside capture against the ruinous cost of large drawdowns.
For a binary bet (win or lose), the formula is:
f* = (b × p − q) / b
# Where:
b = net decimal payout per unit staked # e.g. 100/110 ≈ 0.909 for −110 American
p = your true probability of winning
q = 1 − p # probability of losing
f* = fraction of bankroll to wager
At negative American odds, convert: b = 100 / |odds|. So minus 110 gives b = 100/110 = 0.9091. At positive American odds, b = odds / 100, so plus 120 gives b = 1.20.
The formula derives from taking the derivative of E[ln(W')] = p·ln(1+fb) + q·ln(1-f) with respect to f, setting the derivative to zero, and solving. No assumptions beyond the bet parameters. Pure first-principles optimization of the log-wealth function.
Worked example: 55 percent win rate at minus 110
Take a bettor with a genuine 55 percent win rate on NFL spread bets at standard minus 110 juice. The break-even at minus 110 is 52.38 percent (110 divided by 210), so the bettor holds a real 2.6-point edge over the implied probability.
p = 0.55 # true win rate
q = 0.45
f* = (0.9091 × 0.55 − 0.45) / 0.9091
= (0.5000 − 0.45) / 0.9091
= 0.05 / 0.9091
= 5.50% # full Kelly: bet 5.5% of bankroll
The growth rate per bet, the expected logarithmic return each play, is:
# Full Kelly (f = 5.50%):
G(0.0550) = 0.55 · ln(1.05) + 0.45 · ln(0.945)
= 0.55 × 0.04879 + 0.45 × (−0.05659)
= +0.1369% per bet
At 1,000 bets, full Kelly grows a $1,000 bankroll to:
e0.001369 × 1000 = e1.369 ≈ $3,930
In expectation, yes. The expectation gets dominated by a few lucky paths. The median outcome sits substantially lower.
The growth-rate curve and why overbetting kills you
The growth rate function G(f) is concave, peaks at f = f*, and crosses zero at exactly f = 2f*. Beyond twice the Kelly fraction, expected log-growth turns negative. You will systematically lose money over time even though every individual bet is +EV. This counters intuition, yet the math is hard. Overbetting destroys more in bad runs than wealth gains in good ones.
For the minus 110 / 55 percent example: full Kelly = 5.50 percent, so 2× Kelly = 11 percent. At 11 percent, you bet the maximum before guaranteed long-run ruin. At 6 percent, a half percentage point above full Kelly, you cross into sub-optimal territory with meaningfully lower growth. At 12 percent, you gamble your way to bankruptcy with a positive-edge bet.
The comparison table shows the growth rate, 1,000-bet outcome, and variance at each fraction:
| Kelly fraction | Bet size | Growth/bet | $1k → 1,000 bets | Relative variance | Assessment |
|---|---|---|---|---|---|
| 2× Kelly | 11.0% | ≈ 0% | ≈ $1,000 | 400% | Zero growth, extreme risk |
| 1.5× Kelly | 8.25% | +0.069% | $2,000 | 225% | Worse than half-Kelly, higher variance |
| Full Kelly (1×) | 5.50% | +0.137% | $3,930 | 100% | Max growth, max volatility |
| Half Kelly (0.5×) | 2.75% | +0.103% | $2,800 | 25% | 75% of growth, 75% less variance |
| Quarter Kelly (0.25×) | 1.375% | +0.060% | $1,820 | 6.25% | 44% of growth, 94% less variance |
| Eighth Kelly (0.125×) | 0.6875% | +0.032% | $1,380 | 1.6% | 23% of growth, near flat-staking |
The key asymmetry: variance scales with f2 (the square of the fraction), while growth scales with roughly f·(1−f/2). Going from full Kelly to quarter-Kelly cuts the fraction by 75 percent. The variance ratio works out as (f_Q/f_full)2 = (0.25)2 = 0.0625, so variance drops by 93.75 percent. Growth drops by only 56 percent. The variance reduction lands disproportionately large relative to the growth sacrifice. The asymmetry is the entire mathematical case for fractional Kelly.
The real argument for quarter-Kelly: edge estimation error
The variance argument alone would justify half-Kelly, not quarter. The deeper reason professional bettors use quarter or less is edge estimation error. Your model's predicted edge sits above your true edge almost every time, and full Kelly on an overestimated edge is a much larger problem than the surface math shows.
Consider the scenario: you believe you win 55 percent on NFL spreads at minus 110. You have tracked 200 bets, hit 52 percent (sample variance is high at 200 bets), but your model says 55 percent. You run Kelly, get 5.5 percent, and size accordingly.
What if your true win rate is 53 percent?
f*_true = (0.9091 × 0.53 − 0.47) / 0.9091
= (0.4818 − 0.47) / 0.9091
= 0.0118 / 0.9091
= 1.30% # true Kelly is only 1.3%
# But you're betting 5.5% (your estimated full Kelly):
multiple of true Kelly = 5.50% / 1.30% = 4.23× # overbetting by 4×
You bet 4.23× your true Kelly fraction, well past the 2× threshold where growth goes negative. You hold a genuine edge (53 percent above 52.38 percent break-even), but you bet at a size guaranteeing long-run losses. The math does not care about good intentions.
Now watch what quarter-Kelly does to the same scenario:
multiple of true Kelly = 1.375% / 1.30% = 1.06×
# Close to optimal for your actual edge, by accident
Quarter-Kelly on your overestimated edge lands almost exactly on true Kelly for your real edge. This is the structural reason professional bettors use quarter-Kelly. The fraction provides a natural hedge against the systematic overestimation baked into any model selection process. You do not pick bets you think are minus EV. You pick bets your model says are plus EV. The selection bias means your average estimated edge always sits above your average true edge.
What to use at each stage
The right Kelly fraction depends on how confident you are in your edge estimates and how much sample data you have:
- Under 300 tracked bets: use quarter-Kelly or less. Sample variance is high enough so 55 percent in 200 bets stays consistent with a true rate anywhere from 48 percent to 62 percent. Full Kelly on 200-bet samples gambles on your model, not your edge.
- 300 to 1,000 tracked bets, consistent results: quarter to half-Kelly. Your edge estimate is tightening but still carries material uncertainty. Half-Kelly at this stage is a reasonable ceiling.
- 1,000-plus bets, model validated: half-Kelly as a ceiling. Even sharp professional bettors rarely exceed half-Kelly. The argument for full Kelly only holds if your edge estimate is exact, which never happens.
- Using public consensus, no proprietary model: stay at quarter-Kelly or eighth-Kelly. If your edge comes from the same public information everyone else has, the edge is small and fragile. Size accordingly.
Platforms like Outlier and Pikkit track your bets over time. The long-run CLV (closing line value) data they generate is the best proxy for true edge you have access to without building a proprietary model from scratch. When Pikkit shows you have beaten the closing line by 3 percent over 500 bets, plug the number into Kelly. Not your model's prediction.
Bankroll simulator
The widget below runs 500 simulated bets for each Kelly fraction simultaneously, using the win probability and odds you enter. The three curves show the median outcome (solid line) and the 10th to 90th percentile band. The gap between the bands is the real story. Full Kelly has enormous variance reaching 10× growth on lucky paths and producing 80-plus percent drawdowns on common ones.
The vig math and why your edge is smaller than you think
There is one more layer bettors miss. When you use the standard minus 110 / minus 110 market at a major US book, you are not seeing the true market price. You are seeing a vigged price. The break-even at minus 110 is 52.38 percent (110 divided by 210), not 50 percent. The book has already extracted about 4.5 percent hold from the two-sided market before your bet is placed.
Your edge is calculated against the no-vig fair line, not against the 50/50 coin flip:
implied_A = 110 / 210 = 52.38%
implied_B = 110 / 210 = 52.38%
total_implied = 104.76% # the vig
fair_A = 52.38% / 104.76% = 50.00%
# Your edge on a 55% true-prob bet at −110:
edge = 55% − 52.38% = 2.62 percentage points
When you use a sharp book like DraftKings or compare against Pinnacle closing lines for CLV purposes, you want to de-vig the market first. The implication: if a tool reports your edge as the raw comparison to 50 percent, the tool overstates your edge by roughly 2.5 points on a standard market. Run de-vigged numbers into Kelly. The R library bettoR (papagorgio23) includes implied probability conversion utilities for this step.
Quarter-Kelly is not conservative, the choice is correct
The internet's typical framing of fractional Kelly is wrong. The standard description calls fractional Kelly "conservative Kelly for people who cannot handle volatility", a performance-for-comfort trade-off. The framing misses the deeper point.
Quarter-Kelly is the mathematically correct fraction for a bettor with imperfect edge estimates. The proof is in the estimation-error math. The systematic gap between estimated and true edge means your full estimated Kelly corresponds to roughly 3 to 5× your true Kelly. Quarter of your estimate lands near 0.75 to 1.25× your true Kelly, optimal or near-optimal by definition.
The open-source Kelly library keeks (wdm0006) implements this directly with a DrawdownAdjustedKelly variant. Other bankroll tools use configurable Kelly fractions because practitioners know the full fraction does not survive contact with real data. Expected value is what you size from. Kelly translates the sizing into bankroll fractions, and the fraction you use should reflect your confidence in the edge estimate, not the edge magnitude alone.
The practical workflow: use your best model to identify plus-EV plays. Convert the American odds to net decimal (b), compare your estimated win probability (p) to the no-vig break-even, plug into f* = (bp minus q) divided by b, then apply the fraction appropriate to your sample size and confidence level. For most bettors tracking under 500 bets, quarter-Kelly. Beyond 500 with consistent CLV, consider half. Full Kelly only fits a proprietary model with thousands of resolved bets and a willingness to stomach 80-plus percent drawdown paths on the distribution's tail.
Use FanDuel, BetMGM, and Caesars for the retail books where the boosts and promotions give you the raw edge. Use the math above to size correctly. Edge identification is where most bettors focus. Sizing is where most of the damage happens.