Education
Variance vs Edge: Reading a Drawdown Honestly
A positive edge and a brutal losing streak are not contradictions. Here is how to separate what your model expects from what a single history delivered.
Key takeaways
- Edge grows linearly with the number of bets; noise grows with its square root.
- The break-even horizon — where drift finally clears noise — is roughly (σ / edge)² bets.
- Long losing streaks are expected, not anomalous: at a 50% win rate, ten in a row appears within about a thousand bets.
- Maximum drawdown is a property of one path; read it as a distribution across many seeds.
- Empty statistics should read as undefined, never as a confident zero.
Edge sets the drift; variance sets the ride
Expected value tells you the direction of the drift. Variance tells you how violently the path wanders around it. Over a short sample the wandering dominates, which is why an edge can be real and the equity curve can still spend long stretches underwater.
The two grow at different rates, and that difference is the whole story. After n bets your expected profit is n × edge — linear. The standard deviation of that total is σ × √n — sublinear. So the signal-to-noise ratio improves as √n, slowly and inexorably, and there is a specific n before which the noise is simply larger than the signal.
Set the two equal and solve: drift clears noise at roughly n = (σ / edge)². Flat-staking a 2% edge at even money, with σ near 1 unit per bet, that horizon is about 2,500 bets. Below it, your equity curve is mostly telling you about luck. Above it, it starts telling you about your model.
Losing streaks are not evidence of anything
At a 50% win rate, the probability of a specific run of ten consecutive losses is 1 in 1,024. That sounds rare until you ask the right question, which is not 'how likely is this specific run' but 'how likely am I to see at least one such run somewhere in my season'. Across a thousand bets the answer is better than even money.
Longer streaks follow the same arithmetic. Fifteen straight losses at 50% is a 1-in-32,768 sequence, and across 50,000 bets you should expect to live through one. At a 35% win rate — perfectly normal when you are taking plus-money prices — a run of ten losses appears roughly every 200 bets.
So a streak is not a signal. It is a sample from a distribution you already knew the shape of. The useful question is whether the observed streak length is extreme relative to your win rate and sample size, and that is a calculation, not a feeling.
Max drawdown is a path statistic
Maximum drawdown answers one question: how far did the bankroll fall from a prior peak? Two staking plans with identical end results can have wildly different drawdowns, and the deeper one is more likely to end a real bankroll before the edge pays off.
Read drawdown alongside the longest losing streak and the ruin rate across many paths, not from a single lucky or unlucky run.
The subtlety is that maximum drawdown is a maximum — an extreme-value statistic — and extremes grow with observation length. Run twice as many bets with the same edge and staking plan and your expected maximum drawdown gets worse, even though your expected profit gets better. A drawdown figure quoted without the number of bets it was measured over is not a number you can use.
When something looks too good or too bad
When a result looks too good, check the sample size, the selection share and whether the lab injected a known edge. When it looks too bad, check variance and drawdown before blaming the maths. Empty statistics should read as undefined, not as a confident zero.
A short checklist that resolves most surprises. How many selections, not how many events? Does the confidence interval on the observed rate contain the expected rate? Does the result survive three different seeds? Was a known edge configured into this run? Is the metric averaging over an empty set?
That last one deserves emphasis. A win rate over zero selections is undefined, and reporting it as 0% invites exactly the wrong conclusion. Novus Odds renders empty statistics as undefined for that reason — a missing measurement and a measured zero are different claims about the world, and conflating them is how confident nonsense gets made.
Frequently asked
How many bets before an edge becomes visible?
Roughly (σ / edge)² bets, where σ is the standard deviation of a single unit outcome. Flat-staking a 2% edge at even money puts that horizon near 2,500 bets. Below it the equity curve is dominated by noise.
Is a ten-bet losing streak unusual?
No. At a 50% win rate a specific run of ten has probability 1 in 1,024, but across a thousand bets you are more likely than not to encounter at least one. At the 35% win rates typical of plus-money prices, ten in a row shows up every couple of hundred bets.
Why did my maximum drawdown get worse with more bets?
Because maximum drawdown is an extreme-value statistic and extremes grow with observation length. Longer runs give the path more opportunities to find a deep trough, even while expected profit rises. Always quote drawdown alongside the sample it was measured over.
Why does the lab show undefined instead of zero?
Because a statistic computed over an empty set has no value, and displaying 0% would assert something false. A missing measurement and a measured zero are different claims, and the labs keep them distinct.
Try it yourself
Everything in this article is something you can run and tweak in the lab — with your own settings and a reproducible seed.
Open the lab →