Kelly Criterion Application

The Kelly Criterion answers a single question: if your probability estimate is right, what fraction of bankroll maximises long-run growth? It is powerful, elegant, and dangerous when fed fantasy inputs.

This lesson focuses on application: the formula, fractional Kelly, football worked examples, hard caps, and when to leave Kelly on the shelf.

The Formula (Decimal Odds)

For a bet at decimal odds o with your estimated true probability p (and assuming 0 < p < 1, o > 1):

Net odds multiple b = o − 1
Full Kelly fraction f* = (p × o − 1) / (o − 1)

Equivalently: f* = edge / b, where edge in stake terms is p × o − 1 when positive.

• If f* ≤ 0, Kelly says do not bet (no edge under your p).
• If f* > 0, full Kelly suggests staking fraction f* of bankroll.

Kelly assumes:

1. You may rebet indefinitely with compounding.
2. Your p is calibrated (not optimistic).
3. Odds are available at size for your stake.
4. Outcomes are independent enough for the growth proof to matter over a long series.

Football markets violate these assumptions gently every weekend. That is why fractional Kelly exists.

Worked Example: Premier League Home Favourite

Bankroll: £2,000
Your estimate: Home win p = 0.58
Best price: o = 1.90

Edge term:

p × o − 1 = 0.58 × 1.90 − 1 = 0.102 → +10.2% EV

Full Kelly:

f* = 0.102 / (1.90 − 1) = 0.102 / 0.90 = 0.113 → 11.3% of bankroll

Full Kelly stake ≈ £226.

For most people that is an emotional and practical non-starter on a single football match — even with a genuine edge. One wrong p (say true p was 0.52) flips the story:

0.52 × 1.90 − 1 = −0.012 → negative EV → Kelly stake = 0

Your sizing error was really a calibration error.

Fractional Kelly (What Practitioners Actually Use)

Common policy:

| Policy | Fraction of f* | Stake in example | |---------------|-----------------|------------------| | Full Kelly | 100% | ~£226 | | Half Kelly | 50% | ~£113 | | Quarter Kelly | 25% | ~£57 | | Eighth Kelly | 12.5% | ~£28 |

Quarter Kelly is a widely taught default for noisy sports estimates: you still scale with edge and odds, but variance and model error hurt less.

Then apply a hard unit ceiling (e.g. never more than 2–3% of bankroll on one selection regardless of Kelly). In the example, a 2% cap is £40 — below even quarter Kelly. The cap wins. That is healthy.

Underdog Example

Bankroll: £2,000
Away win p = 0.28
Odds o = 4.20

p × o − 1 = 0.28 × 4.20 − 1 = 0.176
f* = 0.176 / 3.20 ≈ 0.055 → 5.5% full Kelly (~£110)
Quarter Kelly ≈ £27.5 (about 1.4% of bankroll).

Long shots with real edge still get modest stakes under fractional Kelly. If your process wants a flat 1 unit = 1%, you might stake £20 and treat Kelly as a ceiling check rather than a command.

Integrating SupaBola Inputs

SupaBola surfaces model-oriented probabilities and edges on Value Bets, match context on Predictions, coaching narrative on Coach Bola, and deeper cuts on Analytics.

Treat model p as:

• A candidate probability, not gospel.
• Something you may haircut if the market is sharp, news is incomplete, or the league is outside your competence.
• An input to Kelly only after you accept it as your betting probability.

Never run full Kelly on an unexamined dashboard number. That is how “professional formula” becomes amateur liquidation.

Common Mistakes

1. Overstated p. The #1 Kelly killer. Optimistic probabilities explode stake size.
2. Full Kelly on correlated bets. Multiple bets on the same match (or same team across markets) are not independent; naive Kelly overbets the fixture.
3. Ignoring limits. If the book only accepts £15, Kelly’s £80 suggestion is academic.
4. Recalculating after every emotion. Update bankroll on a schedule (e.g. weekly), not after each lost accumulator leg.
5. Using Kelly to justify oversized “confidence” bets. If you need Kelly theatre to stake 10%, you are not managing risk — you are negotiating with dopamine.
6. Applying Kelly to negative EV “opinions”. If edge ≤ 0, the correct Kelly stake is zero. Walk away.

Practical Policy Template

Copy and adapt:

1. Qualify the bet: positive EV after honest p and vig-aware pricing.
2. Compute full Kelly f* from decimal odds.
3. Take k × f* with k ∈ 0.5 (start at 0.25).
4. Apply min( that stake, unit cap, book max ).
5. Log p, o, f*, final stake, reason for haircut.
6. Review calibration monthly: are your 60% picks winning near 60%?

This policy survives bad weeks better than pure full Kelly.

When Not to Use Kelly

• Bankroll is money you cannot afford to vary. Use flat fractional units or do not bet.
• Sample of edges is tiny. Without calibration history, p is storytelling.
• Promotional or free-bet mechanics. Boosted odds and free bets need separate staking logic.
• Live markets moving every few seconds. By the time Kelly is computed, the price may be gone; pre-set rules beat live calculus.
• You are still learning the sport model. Flat 1% stakes teach process; Kelly amplifies whatever process quality you already have — good or bad.
• Portfolio of many simultaneous bets. Simple single-bet Kelly ignores interdependencies; advanced portfolio Kelly is out of scope here and easy to get wrong.

Relating Tools Across This Module

1. Vig mastery — clean market probabilities so you know what you are beating.
2. Odds conversion — feed the correct decimal o into Kelly.
3. Arb basics — when prices alone lock profit, Kelly is irrelevant; stake the arb weights instead.
4. Kelly — when you believe p beats the price, size the bet without blowing up.

Professionals are boring on purpose: small edges, modest fractions, ruthless record-keeping. Kelly is a scalpel, not a slot-machine lever.

Closing the Module

Professional betting tools do not remove risk. They remove unnecessary confusion: margin, formats, false sure-things, and reckless stake sizing. Combined with bankroll discipline and honest evaluation, they support a process you can audit.

If a tool’s output cannot be explained in one paragraph — price, probability, stake, risk — do not use it until you can.

Key Takeaways

• Full Kelly: f* = (p × o − 1) / (o − 1) with decimal odds o and probability p; bet only if f* > 0.
• Prefer quarter or half Kelly, then impose a hard % cap per bet.
• Kelly multiplies calibration error — bad p is more dangerous than no formula.
• Use SupaBola model outputs on Value Bets and Coach Bola as inputs to judgment, not auto-pilot stake commands.
• When assumptions fail (correlated bets, limits, no edge), do not force Kelly.

For educational and informational purposes only. Gambling involves risk. Please bet responsibly.