Variance and sample size in betting

A good process can still produce an ugly run of results, and that cuts both ways for winners and losers. The rest of this page is about why that happens, and what actually changes as more bets get added to the pile.

What variance means in betting

Every bet has two things attached to it: the result you get and the result you would expect on average if you could replay that exact bet thousands of times. Variance is the gap between those two. It is why a good decision can still lose and a bad decision can still win. It is not a flaw in a betting strategy, it is a basic feature of any process built on real uncertainty, and football has plenty of that, with 90 minutes of chance events layered onto two teams’ actual quality.

A bettor who has worked out a genuine small edge, for example backing prices that sit a little better than the true probability, is still exposed to variance on every single bet. The edge describes the average outcome over a long run. It says nothing about what happens on any one bet, or any short run of them. See what is expected value for the full explanation of that average, and why a positive EV bet can still lose.

A losing run doesn’t mean the edge is gone, and a winning run doesn’t prove one exists

Picture a bettor with a small but real edge, similar to a coin very slightly weighted in their favour instead of landing 50/50. Over enough flips, that coin lands in their favour more often than not. Over 50 or 100 flips, though, an unlucky run is entirely possible. The coin does not know it is supposed to average out yet.

The same applies to a bettor placing 50 or 100 bets with a small real edge. Losing runs of several bets in a row happen regularly even to someone who wins more often than they lose overall, purely because a small edge is a small tilt on a result still mostly decided by chance. A losing stretch across 50 or 100 bets sits well within normal variance for someone with a real, modest edge. It does not prove the edge has disappeared, and treating it as proof is one of the most common mistakes in judging a betting record.

The same logic runs the other way. A bettor with no edge at all, or even a small negative edge, can string together a winning month or a winning year. A short run of wins feels like proof of skill, but a coin with no bias at all still lands on the same side several times running sometimes too. A small sample cannot reliably tell a skilled bettor from a lucky one in either direction.

Why sample size is what actually separates skill from luck

The reason a bigger sample matters comes down to how variance and edge behave differently as bets pile up. Standard deviation, the standard way to measure the size of likely swings around an average result, grows with the square root of the number of bets placed. The expected profit from a real edge grows in direct proportion to the number of bets instead. Wikipedia’s summary of the underlying gambling mathematics puts it plainly: “the standard deviation is proportional to the square root of the number of rounds played, while the expected loss is proportional to the number of rounds played,” and because of that gap, the expected result “increases at a much faster rate” than the swings around it as the number of rounds grows (Wikipedia, gambling mathematics).

In plain terms: after a handful of bets, ordinary variance can be as large as, or larger than, any real edge, which is exactly why a small sample tells you almost nothing on its own. As the sample grows, the edge (or the lack of one) keeps accumulating steadily while the swings around it grow more slowly, so eventually the edge becomes visible above the noise. That’s the law of large numbers doing its work, not a guarantee that arrives at any particular bet count, but a trend that keeps strengthening the more bets get added.

A small sample can hide a real edge just as easily as it can fake one. More bets are the only thing that reliably tells the two apart.

How many bets that takes in practice depends on the size of the edge and the odds involved, and different sources give different rules of thumb rather than one fixed number. Working through the maths for a bettor with a 5% yield at typical odds of 2.0, RebelBetting found that after 100 bets there was still roughly a one-in-three chance the result was down to luck alone, and that it took around 1,100 bets at those same parameters before the result reached the commonly used 95% statistical confidence threshold (RebelBetting). That is one worked illustration tied to a specific edge and specific odds, not a universal rule. A bettor with a bigger edge, or different odds, would reach a confident answer after a different number of bets. The takeaway is not a bet count to memorise. It is the general shape: samples of a few dozen or a few hundred bets are rarely enough on their own, however tempting they are to read as proof.

Closing line value gets there faster

Raw profit and loss is a slow way to separate skill from luck precisely because it carries all the variance of the actual results with it. Closing line value strips a lot of that noise out by comparing the price a bettor got against the market’s own final price for the same game, rather than against a random result. See what is closing line value for the full explanation, but the part relevant here is that CLV needs a smaller sample to become meaningful than plain win/loss record does, because it measures pricing skill directly instead of waiting for that skill to show up in results after variance has had its say.

That does not make CLV instant either. A handful of bets with strong CLV is still a small sample. It just reaches a reliable answer faster than results alone, which is why serious bettors track it as an early signal rather than waiting on a full season of profit and loss.

Judging any short record

Put together, this is why a tipster’s, a model’s or a bettor’s short losing streak, or short winning streak, isn’t reliable evidence on its own. A losing month can sit on top of a genuinely good process. A winning month can sit on top of a genuinely bad one. Neither proves anything by itself, and judging either as if it settles the question is judging noise, not skill.

The honest way to read a betting record is to ask how big the sample is before drawing a conclusion from it, weigh closing line value alongside raw results because it converges faster, and accept that even a real, well-founded edge will sometimes produce results that look bad for a while. That discomfort isn’t a sign something has gone wrong. It’s what variance looks like from the inside.