Variance is the main reason strong betting processes can look broken over short samples. A positive edge does not remove drawdowns — it only improves long-run expectation. This guide explains the maths of variance, the distinction between expected value and observed outcomes, and the practical implications for bet sizing, drawdown tolerance, and process evaluation. Read it alongside bankroll management and Kelly for the operational framing.
What variance is
Expected value (EV) is the long-run average result per bet. Variance is the dispersion of outcomes around that expectation. Both are always present in betting — you cannot have EV without variance, and the higher the EV, the more meaningful variance becomes as a tool for separating signal from noise.
Formally, variance is the average squared deviation of outcomes from the mean. Standard deviation (the square root of variance) is more practically useful because it is in the same units as the outcome itself. For a $100 bet at $1.91 odds with 53% win probability:
- Expected value per bet: $1.23 (0.53 × $91 - 0.47 × $100)
- Standard deviation per bet: ~$95
- EV-to-SD ratio: ~0.013 — your edge is tiny relative to single-bet variance
That ratio is why hundreds of bets are needed to reliably detect edge.
Why variance dominates short samples
Over 10 bets, even a strong process can lose 30% of bankroll. Over 100 bets, drawdowns of 15-20% are common despite positive long-run EV. Over 1,000 bets, the law of large numbers starts to take over and realised P&L converges toward expected P&L.
Most punters quit too early. The mental error is judging strategy from realised outcomes rather than process quality. A losing month means almost nothing; a year of consistently negative closing line value means a lot.
Variance by strategy type
- Arbitrage: Near-zero variance per bet because both sides are covered. Total volatility comes from execution errors and balance reconciliation.
- Main-market value betting: Moderate variance. EV is ~2-4% per bet, SD is ~95-100% of stake.
- Props and long-odds strategies: Higher variance because outcomes are longer-tailed. A 30% hit rate at +250 odds has high SD per bet despite positive EV.
- Middling: Very lumpy. Most middle bets resolve as small losses (-5%) with rare large wins (+95%). Requires many attempts before realised matches expected.
- Futures: Extreme variance. Single positions resolve months later with binary outcomes. Bankroll volatility is dominated by 1-2 positions.
- Live/in-play: Variance depends on what bet types you use. Match winner switches in-play have moderate variance; prop overlays in-play have higher.
Bankroll volatility maths
For independent bets with equal stake, total variance after N bets is N times the per-bet variance. Standard deviation scales with the square root of N.
Practical implication: doubling your bet count doesn't double your SD — it multiplies it by 1.41. Your edge (EV) scales linearly, but your noise scales as a square root. That is why volume eventually wins: linear growth of edge outruns square-root growth of variance.
With unit stake u, edge e (as fraction of stake), per-bet SD s, and N bets:
- Expected profit: N × u × e
- SD of total profit: u × s × sqrt(N)
- Sharpe ratio (signal-to-noise): e × sqrt(N) / s
With e = 0.03 (3% edge), s = 0.95, and N = 1000, Sharpe is ~1.0 — the kind of number where realised P&L starts to look like expected. With N = 100, Sharpe is ~0.32 — random variation will dominate.
Risk of ruin and drawdown probabilities
Risk of ruin (RoR) is the probability your bankroll hits zero before reaching some target. RoR depends on edge size, variance per bet, unit size as fraction of bankroll, and the bankroll target.
Rough rules of thumb for a +EV bettor with 2-3% edge using 1% units:
- 10% drawdown: probability ~50% within 200 bets
- 20% drawdown: probability ~25% within 500 bets
- 30% drawdown: probability ~10% within 1000 bets
- 50% drawdown: probability ~2% within 5000 bets
These probabilities are not "if I'm unlucky" — they are the normal range of outcomes for a profitable bettor. Plan capital so you can absorb them without forced position closure or operational stress.
How long edge takes to appear
Most processes need 500-1000 bets for reliable performance signal. Below that, random variation is often larger than edge. The exact number depends on your edge size and bet variance.
Quick guide: bets needed to detect edge with 95% confidence ≈ (2 × s / e)². For e = 0.03 and s = 0.95, that's ~4000 bets. For e = 0.05, it's ~1400 bets. Higher-edge strategies confirm themselves faster.
Most Australian punters never bet enough volume to confirm edge from realised P&L alone. That is why closing line value is the preferred early diagnostic — it stabilises in 100-200 bets rather than thousands.
CLV vs realised P&L as diagnostic
Realised P&L over short windows tells you almost nothing about strategy quality. Closing line value (CLV) is much more informative because it measures process quality directly against the sharp market consensus.
Positive CLV + negative P&L over 100 bets = bad luck. Keep going.
Negative CLV + positive P&L over 100 bets = good luck on a bad process. Reduce stakes and audit the model.
See closing line value explainedfor measurement methodology.
Sizing and drawdown control
Fractional Kelly reduces volatility meaningfully compared with full Kelly. Quarter Kelly is a common practical baseline for Australian advantage bettors balancing growth and drawdown tolerance.
Full Kelly produces optimal log-bankroll growth but with extreme drawdowns. Quarter Kelly captures ~94% of full Kelly's growth at ~25% of its volatility. Half Kelly is the middle ground.
Flat staking (constant unit size, ignoring Kelly) is a reasonable default for bettors without precise edge estimates. Use 1-2% of bankroll per bet on main markets, 0.5-1% on props, 0.25-0.5% on highly variable derivatives.
Never increase stake size during winning streaks. Variance produces streaks naturally — an upward streak is not evidence of higher edge.
Variance and psychology
The psychological cost of variance is the main reason +EV bettors fail in practice. Even knowing intellectually that a 20-bet losing streak is normal does not prevent the urge to scale stakes, switch strategies, or quit.
Mitigations:
- Pre-commit to stake sizing and only review monthly.
- Track CLV alongside P&L — when P&L is negative but CLV is positive, you have objective evidence the process is working.
- Compare your drawdown to simulated drawdown distributions for your edge size.
- Reduce session length and bet frequency during emotional periods.
Common variance mistakes
- Quitting during normal drawdowns. A 20-unit drawdown is well within the expected range for any profitable bettor.
- Scaling stakes during hot streaks. Streaks are variance, not skill. Scaling captures the noise, not the signal.
- Judging strategy quality from tiny samples. Less than 200 bets is not enough for meaningful inference.
- Confusing high frequency with low variance. Many low-stakes bets still produce significant cumulative variance.
- Ignoring correlation. Multiple bets on the same match are not independent; their variance compounds faster than uncorrelated bets.
- Chasing losses with bigger stakes. Increases drawdown depth without improving expected recovery.
- No CLV log. Without CLV you cannot distinguish bad luck from bad process.
Operations checklist
- Log every bet with stake, price, closing line, closing price.
- Calculate rolling CLV (last 50, 100, 200 bets).
- Track drawdown depth from running peak.
- Simulate expected drawdown distribution for your edge/SD/N.
- Use fractional Kelly (0.25-0.5) not full Kelly.
- Never increase stake size during streaks (up or down).
- Review process monthly, not weekly.