Sample Size and Statistical Significance

The number nobody tells beginners

To prove with 95% confidence that you're a true 55% bettor at -110 odds, you need around 1,200 to 1,400 settled bets. To prove with 95% confidence that you have a 4% ROI edge, you need around 5,000 to 7,000 bets. These aren't ballpark guesses from somebody on Twitter. They fall straight out of the binomial standard error and the central limit theorem.

Most retail bettors place 200 to 500 bets a year. At that rate, proving an edge to yourself with real statistical confidence takes between three and ten years. That's the structural reason most bettors who think they have an edge are wrong about it. It's also the reason that bettors who genuinely do have an edge can spend years unable to tell themselves apart from the unlucky-no-edge group on the data alone.

Setting up the math

Each bet is a Bernoulli trial. It wins with probability p or loses with probability 1 − p. Across n bets, the number of wins follows a binomial distribution with mean np and variance np(1 − p). Win rate is the sample proportion p̂ = wins / n, with standard error:

For large n, the sample win rate is approximately normally distributed around the true win rate, so a 95% confidence interval is roughly p̂ ± 1.96 · SE(p̂).

Worked example. Same bettor as the subtitle: 110-90 over 200 -110 bets. Sample win rate is 0.55. Standard error using the sample value:

So this bettor's 95% confidence interval for true win rate runs from 48.1% to 61.9%. The breakeven win rate at -110 is 52.38%. The interval contains 50% (no edge, just lucky) and 60% (massive edge, professional). The data is genuinely consistent with both. After 200 bets you cannot tell whether you're a coin flip on a hot streak or a serious edge bettor having a slow start.

The breakeven win rate at any price

To convert American odds into a breakeven win rate (the win rate where the bet has zero EV before considering vig elsewhere):

Beating breakeven by N percentage points produces an EV per bet that depends on the price. Beating -110 by 2 points (winning 54.38% vs the 52.38% breakeven) gets you about 3.8% EV per dollar wagered. Beating +300 by the same 2 points (27% vs 25% breakeven) gets you about 8% EV per dollar wagered. Beating breakeven by the same percentage on longer prices is worth more per bet, but it takes a bigger sample to confirm.

Sample size for a stated win rate edge

To detect a true win rate of p as significantly different from breakeven b at a confidence level z, the required sample size is approximately:

For a true 55% win rate at -110 (b = 0.5238):

For a true 54% win rate at -110 (smaller, more realistic edge):

For a true 53% win rate at -110 (small but still profitable):

The non-linearity here is brutal. A 55% bettor needs around 1,400 bets to prove themselves. A 53% bettor (still solidly profitable) needs nearly 25,000 bets. Smaller edges require dramatically larger samples. Bettors with marginal edges effectively cannot prove their edge to themselves within a betting career, even if the edge is real.

Standard error of ROI

Win rate works fine when every bet is the same price. Most real betting samples involve a mix of prices, so the more general statistic is ROI per bet, computed across every settled wager regardless of price. The math gets heavier.

For a single bet at decimal odds o with stake 1 and true win probability p, the per-bet return R is:

Variance of return per bet at decimal odds o with true win probability p is p(1 − p)o2. Clean result. Variance scales with o2, so longshots have wildly more variance per bet than favorites. A +500 bet (decimal 6.0) has roughly 36 times the per-bet variance of a -110 bet (decimal 1.91, so o² ≈ 3.65) at equivalent stakes.

For a bettor placing N bets at varying prices, the variance of average ROI is the average per-bet variance divided by N:

Plug in numbers for a bettor placing 1,000 -110 bets (decimal 1.91) with a true win rate of 0.5238 (the breakeven, so true edge is exactly zero):

A no-edge bettor placing 1,000 -110 bets has a 95% chance of finishing somewhere between −5.92% ROI and +5.92% ROI on luck alone. So a bettor finishing the year at +4% ROI cannot statistically distinguish themselves from a true zero-edge bettor at all. The same bettor finishing at +7% ROI just barely clears the 95% bar, and only if their sample is exactly 1,000 bets.

Required sample size for a stated ROI edge

To prove a true ROI of r at confidence level z, the sample size you need is roughly:

For a true 4% ROI edge at uniform -110 pricing:

For a true 2% ROI edge:

For a true 1% ROI edge:

Same pattern as with win rate, just expressed in the metric most bettors actually track. A 4% ROI edge takes around 2,200 bets to confirm. A 1% ROI edge takes 35,000. There's a real reason serious bettors are obsessive about line shopping for an extra 1% to 2% of EV: that extra bit of edge changes the sample size required to prove the edge from a lifetime down to a year.

Reverse calculation: what does my sample actually tell me?

The more useful question for most bettors is the inverse. Given N bets at observed ROI r̂, what's the 95% confidence interval for true ROI?

For uniform -110 pricing, you can approximate s as 0.95. So the rule of thumb for a -110 bettor:

A bettor at +5% ROI over 1,000 bets at -110 cannot reject the null hypothesis that their true edge is zero. The lower bound of the 95% CI is −0.9%, which still includes zero. They have to push the sample to 1,500 or 2,000 bets to clear that bar.

Why CLV is the smarter signal

Closing line value collapses the sample-size problem. CLV is asking a different question: did the bet you placed beat the closing line? That comparison is direct, sport-independent, and gives you a usable signal in dozens of bets instead of thousands.

The mechanism: the closing line is the most informed estimate of the true probability available before the game starts. If your bets consistently beat the closing line, you're getting better odds than the market consensus, which mathematically has to produce profitable betting if you keep doing it. CLV is a leading indicator, not a confirming one.

Variance of CLV per bet is much smaller than variance of bet outcomes. CLV is the difference between two market estimates of the same probability, not the realization of the underlying random outcome. So the standard error of average CLV is much smaller, and the sample size you need for confidence is around 50 to 100 bets, not thousands.

A bettor consistently beating the close by 2% is showing a real edge. They don't need to wait for results to confirm it. On the flip side, a bettor with +5% ROI over 200 bets but flat or negative CLV is almost certainly running hot. The results will mean-revert downward.

For working with CLV directly, see the dedicated CLV guide.

Practical takeaways

• Treat any conclusion drawn from fewer than 500 bets as basically uninformative. Variance is too high.
• Treat 500 to 1,500 bets as suggestive but not conclusive. Hot and cold streaks within that range are normal and consistent with a wide range of true edges.
• Treat 1,500 to 5,000 bets as the sample where moderate edges (3% to 5% ROI) become statistically distinguishable from zero.
• Smaller edges (1% to 2% ROI) need samples in the tens of thousands and are essentially unconfirmable for most retail bettors over reasonable time horizons.
• If you need a faster answer, track CLV. CLV converges fast enough to give you a directional read in a few months.
• Variance per bet scales with the square of decimal odds. A bettor playing longshots will swing wildly even with the same true edge as a bettor playing favorites. Adjust expectations and sample sizes accordingly.
• The bettor at +5% ROI through 200 bets is in a probability cloud that includes pros, lucky amateurs, and unlucky pros. The data alone can't tell you which one. Ten more years of betting will, eventually.

The honest summary

Most retail bettors who think they're winning at sports are working with sample sizes too small to support that belief. That isn't a moral failing or a math insult. It's a structural feature of low-edge, high-variance gambling. The signal-to-noise ratio is bad enough that even genuinely profitable bettors need years of data to prove their edge to themselves with real statistical confidence.

Same math means running cold for a year, even badly, doesn't necessarily mean your edge is gone. It might mean you lost the variance lottery for that year. Telling real edge erosion apart from variance is one of the genuinely hard problems in retail betting, and the only durable answer is to track CLV alongside results, because CLV is observable in real time and converges fast enough to be useful.

If this guide is depressing, that's part of the point. The math is what it is. The bettors who survive long-term are the ones who size correctly, track honestly, and avoid drawing big conclusions from small samples in either direction.