Kelly fraction in practice — what 0.25× actually does to your bankroll
Full Kelly is mathematically optimal and operationally suicidal. 0.25× Kelly trades ~50% of long-run growth for ~94% less variance per bet — the math behind why every professional bankroll lands here.
- Kelly
- Bankroll
Full Kelly maximizes long-run log-growth of a bankroll, but the variance is brutal: 50%+ drawdowns are common even on a real edge. Fractional Kelly at 0.25× preserves roughly half the expected log-growth while cutting per-bet variance by roughly 94%. That’s the trade every professional bankroll lands on — and the trade Glitch Edge defaults to in code, with a 2% bankroll cap on top.
The Kelly formula, briefly
For a bet at decimal odds o with model probability p:
Kelly fraction f* = (p × o − 1) / (o − 1)
The result is the fraction of bankroll to stake on a single bet. A bet at 2.00 decimal odds with 55% model probability gives f* = 0.10 — full Kelly says stake 10% of bankroll on this bet.
Why nobody actually runs full Kelly
The expected log-growth at full Kelly is the theoretical maximum. The variance around that growth is also the theoretical maximum, and the path is brutal. Three concrete numbers:
- Drawdown distribution. At full Kelly with a real 2% edge, the median drawdown over 1,000 bets is 35–45%. The 95th percentile drawdown is 70%+.
- Recovery time. A 50% drawdown requires a 100% rebound to recover. Most operators emotionally tap out before the rebound; the strategy “fails” not because the edge was wrong but because the human couldn’t hold position.
- Model error compounds. If your model overestimates edge by 50% (common — your real edge is 1.3% not 2%), full Kelly sizes correctly for the estimated edge and dangerously for the actual edge.
Full Kelly is the right answer if you have perfect probability estimates and infinite emotional bandwidth. Nobody does.
What 0.25× actually trades
The math is asymmetric in the fractional direction. At fraction k of full Kelly:
- Expected log-growth: roughly 2k − k² of optimal (so 0.25× gives ~0.44 of optimal log-growth)
- Variance per bet: scales as k² (so 0.25× gives ~6% of full variance)
In plain English: cutting Kelly by 75% costs you roughly half the long-run growth but cuts variance by 94%. That’s a deal almost everyone takes.
| Fraction | Expected log-growth | Variance | Typical max drawdown |
|---|---|---|---|
| 1.0× (full) | 1.00 (optimal) | 1.00 | 60–80% |
| 0.5× | 0.75 | 0.25 | 25–35% |
| 0.25× (standard) | 0.44 | 0.0625 | 10–18% |
| 0.10× | 0.19 | 0.01 | 4–8% |
Why 0.25× and not 0.5×
0.5× Kelly is fine for operators with strong edge confidence and bankroll resilience. The reason most settle at 0.25× isn’t variance — it’s edge uncertainty. Real models are wrong by 20–50% on probability estimates. 0.25× Kelly is conservative enough that even with 50% overestimation, the actual variance you experience is still tolerable.
Why not 0.1× — leaves growth on the table
0.1× Kelly gives only 19% of optimal log-growth. The variance reduction over 0.25× is real but marginal at this point, and you’re sacrificing growth meaningfully. Operators who run 0.1× usually have a specific reason: very high edge uncertainty, regulatory cap on max-per-bet stake, or a thin bankroll where any drawdown matters.
The hard cap on top
Even at 0.25× Kelly, a model overestimating edge by 5× would size 5× larger than appropriate. The second guard is a max-per-bet bankroll cap — typically 2% of bankroll. The cap clips Kelly’s recommendation regardless of edge estimate.
Glitch Edge’s default config:
- Fractional Kelly = 0.25×
- Max per-bet = 2% of bankroll
- Max correlated exposure = 6% of bankroll (sum of bets that could resolve together)
- Max drawdown halt = 25% (strategy halts; operator must un-halt manually)
The platform refuses Kelly fractions above 1× regardless of operator config — full Kelly is too dangerous to ship as a configurable knob.
What this looks like in P/L
A $10,000 bankroll running 0.25× Kelly with a real +2% edge, ~3 bets/day:
- Expected growth: ~6% per year
- 1-year median drawdown: 8–12%
- 1-year 95th-percentile drawdown: 18–22%
- Recovery from median drawdown: ~3 months
A $10,000 bankroll running full Kelly with the same edge:
- Expected growth: ~14% per year
- 1-year median drawdown: 35–45%
- 1-year 95th-percentile drawdown: 65%+
- Recovery from median drawdown: 9+ months
The 0.25× operator gets less growth but survives. The full Kelly operator might get higher growth or might tap out emotionally during the drawdown. Across a population of operators, the 0.25× cohort survives at much higher rates.
What’s NOT fixed by fractional Kelly
Bankroll discipline is a separate concept from bet sizing. Fractional Kelly is the size each bet at this fraction of full Kelly rule. It doesn’t prevent:
- Pressing during a hot streak (operator overrides the fraction emotionally)
- Pressing during a cold streak (“I need to win it back”)
- Adding correlated bets that breach total exposure
- Continuing to bet a strategy after it’s drawdown-halted
These need separate guardrails — max correlated exposure, drawdown halt, no-override policy. The fraction is one of several discipline levers, not the only one.
Frequently asked questions
How do I pick the right Kelly fraction for my situation?
Start at 0.25× unless you have a specific reason to deviate. Move to 0.5× only after 200+ bets of validated CLV. Move to 0.1× if you have very high edge uncertainty or a thin bankroll where any drawdown materially affects you.
Does Glitch Edge enforce my Kelly fraction in code?
Yes. The configured fraction is applied per bet; the platform refuses to size above 1× regardless of operator config. The max-per-bet bankroll cap clips on top.
What if my model probability is wrong?
That’s the main risk. Fractional Kelly + hard cap together hedge against this. If you suspect your model overestimates by ~30%, run 0.25× as if you’re at 0.5× — i.e., your real position is more conservative than the math suggests.
Can I use this for arbitrage or value-bet alerts from other tools?
Yes — Kelly works on any +EV bet regardless of source. Whether the edge came from an OddsJam alert, a RebelBetting value-bet, or your own model, the sizing math is the same. See OddsJam alternative for how Glitch Edge layers under those tools.
What if I want to recover from a drawdown faster?
You don’t. Pressing during a drawdown (“I need to win it back”) is the most reliable way to blow up a bankroll. The discipline is to stay at the same fraction and let the math work over time.