The Kelly Criterion is a formula for sizing bets as a fraction of your bankroll, based on two inputs: your estimated win probability and the payout odds.
Kelly does not tell you what to bet. It tells you how much to bet — and only if you have a real edge.
If your estimated probability doesn't beat the implied probability in the odds, Kelly tells you to bet nothing. That is not a limitation — that is the point.
What Kelly is trying to solve
Most bettors have a staking problem, not just a picking problem.
Two things are simultaneously true:
- Even good bets lose sometimes
- Overbetting kills a bankroll faster than bad picks
Kelly is designed for repeated betting scenarios, where you want long-run bankroll growth without running the obvious failure mode of going broke. It is growth-optimizing, not safety-first. Understanding that difference matters for how you use it.
The basic Kelly formula
For a standard win/lose bet:
f = p − (q ÷ b)
Where:
- f = the fraction of bankroll to wager
- p = your estimated probability of winning
- q = 1 − p (your estimated probability of losing)
- b = net profit per 1 unit risked (not including your returned stake)
If f ≤ 0, the bet is not positive EV at those odds. Kelly's correct answer is "no bet."
Getting b from American odds
| Odds format | How to get b |
|---|---|
| Decimal odds D | b = D − 1 |
| American +A (e.g. +150) | b = A ÷ 100 |
| American −A (e.g. −120) | b = 100 ÷ A |
The inputs you actually need
Kelly only makes sense if you have:
1. A defined bankroll A bankroll is money set aside exclusively for betting. Money you need for living expenses is not a bankroll. Kelly sizing without a real bankroll produces meaningless fractions.
2. A probability estimate based on a repeatable process "I like this side" is not a probability estimate. Kelly requires a number that reflects real modeling, pricing discipline, or a validated track record. If your probability is a guess, Kelly will just scale your mistakes.
3. A price that is meaningfully mispriced If your estimate is not better than the market's implied probability, Kelly produces zero or negative fractions. No formula can create edge where none exists.
→ Implied Probability and How to Calculate It → Expected Value in Sports Betting
How Kelly connects to implied probability and EV
Step 1: Implied probability is your baseline
Odds imply a break-even win rate. Your estimate must exceed it for the bet to have positive EV — and only positive EV bets produce positive Kelly fractions.
Step 2: EV is the gatekeeper
For a 1-unit stake at decimal odds D:
EV = p(D − 1) − (1 − p)
If EV ≤ 0, Kelly produces f ≤ 0.
Step 3: Kelly converts EV into a bankroll fraction
For a binary bet at decimal odds D, Kelly can be rewritten as:
f = (pD − 1) ÷ (D − 1)
The numerator (pD − 1) is the EV per unit staked. The denominator scales by payout. This shows exactly how Kelly and EV relate.
Practical examples
Example 1: Underdog at +150, edge of ~4%
- Odds: +150 → b = 1.50
- Estimated win probability: p = 0.45, q = 0.55
f = 0.45 − (0.55 ÷ 1.50) = 0.45 − 0.3667 = **0.0833**
| Kelly fraction | Bet size on $1,000 bankroll |
|---|---|
| Full Kelly (8.3%) | $83 |
| Half Kelly (4.2%) | $42 |
| Quarter Kelly (2.1%) | $21 |
Those numbers may feel large — especially full Kelly. That reaction is worth respecting.
Example 2: Favorite at −120, small edge
- Odds: −120 → b = 100 ÷ 120 = 0.8333
- Estimated win probability: p = 0.58, q = 0.42
f = 0.58 − (0.42 ÷ 0.8333) = 0.58 − 0.5040 = **0.0760**
Full Kelly: 7.6% of bankroll. Small errors in p swing this dramatically.
Example 3: The most important example — probability error
At −110: b = 100 ÷ 110 = 0.9091
If you estimate p = 0.56:
f = 0.56 − (0.44 ÷ 0.9091) = 0.56 − 0.4840 = 0.0760 → Bet 7.6%
If the true probability is actually p = 0.52:
f = 0.52 − (0.48 ÷ 0.9091) = 0.52 − 0.5280 = −0.0080 → No bet
A 4-point overestimate in your win probability turns a 7.6% Kelly bet into a "no bet" situation. This is why Kelly is only as good as the probability estimate going into it.
Full Kelly vs. fractional Kelly
Why full Kelly is often too aggressive for sports betting
Full Kelly is the mathematically optimal fraction for long-run growth under ideal conditions. In sports betting, the conditions are rarely ideal:
- Your probability estimate is uncertain
- Your edge fluctuates by matchup, market, and timing
- You face practical constraints: limits, price changes, emotional pressure
- A string of losses at full Kelly sizing can be psychologically brutal
Fractional Kelly is the practical default
Fractional Kelly means betting a fixed fraction of the calculated Kelly amount:
- Half Kelly: 0.5 × f
- Quarter Kelly: 0.25 × f
- 10th Kelly: 0.10 × f
Benefits:
- Reduces volatility and drawdown risk
- Provides a buffer against being wrong about your true probability
- Makes bankroll survival over a real season more realistic
For most sports bettors without a validated, calibrated model, quarter Kelly or less — with a hard maximum cap — is the appropriate starting point.
→ Bankroll Management: Unit Sizing and Variance
Kelly vs. flat betting
Flat betting: Fixed stake per bet regardless of edge. Simple, easy to track, hard to overbet.
Kelly: Dynamic sizing — bigger stakes when edge is larger, smaller when edge is thin.
In theory, Kelly grows bankroll faster when edges are real and probability estimates are accurate. In practice, flat betting often outperforms bad Kelly usage because it's much harder to materially overbet. If you are not calibrated — if you don't know your actual hit rate by market and matchup type — flat betting with strict unit rules is safer and still serious.
The case for Kelly only becomes compelling once you have a validated, well-calibrated edge estimate across a large sample.
What Kelly can and cannot be used for
| Market type | Kelly fit |
|---|---|
| Moneyline, spread, total (binary outcomes) | Direct fit |
| Props with binary settlement | Direct fit |
| Parlays (independent legs) | Possible with joint probability estimate |
| Correlated parlays / same-game parlays | Poor fit without a joint model |
| Futures (multi-way) | Requires additional framework |
Common mistakes and guardrails
Mistakes:
- Using Kelly as a system without an actual probability model
- Confusing implied probability with true probability (the market's price is not your estimate)
- Using full Kelly with noisy or unvalidated probability estimates
- Ignoring correlation across simultaneous bets
- Adjusting your staking policy based on a short hot streak
Guardrails:
- Use fractional Kelly (quarter or less is a safe starting point)
- Apply a hard max bet cap per wager (e.g., never more than 3–5% of bankroll)
- Require a minimum edge before betting at all (e.g., your estimate must beat implied probability by at least X%)
- Track closing line movement against your entries — it provides feedback on whether your prices are actually better than the market
- Treat your bankroll as a season-long resource, not a daily swing
Frequently Asked Questions
Is Kelly Criterion good for sports betting? It can be — if you have a probability estimate grounded in real modeling or a validated track record, and you use conservative fractional sizing. Without that, Kelly just scales your uncertainty.
What if Kelly says to bet 0? That means your estimate doesn't beat the break-even implied by the price, or your edge is too small to justify risking bankroll. This is a feature — not a signal to find a different formula.
Should I ever use full Kelly? For most sports bettors, full Kelly is too aggressive because probability estimates in real markets carry meaningful error. Fractional Kelly is the practical solution.
Can I use Kelly for parlays? Only if you can estimate the true joint probability of the parlay outcome, including any correlation between legs. If you can't model correlation, Kelly sizing will be wrong.
How do I find my win probability (p)? Kelly doesn't supply p. You need a process: modeling, pricing, and validation across a large-enough sample to measure calibration. If you're estimating from intuition rather than tracked results, assume your p is noisier than you think — and size down.
CapperU summary
- Kelly is a bet sizing tool, not a pick tool
- It converts a real edge into a bankroll fraction
- The biggest practical risk is not variance — it's being wrong about your probability estimate and betting too large
- Fractional Kelly with a hard cap is the real-world implementation for disciplined sports bettors
- Without a validated edge, Kelly amplifies mistakes rather than correcting them
Put it to use
Apply Kelly sizing discipline to real model outputs with captured odds:
Continue learning: Expected Value in Sports Betting · Implied Probability · Bankroll Management · Closing Line Value