Sample Size and Variance: How Many Bets Before You Know a Model Works

Here is a question worth sitting with. A betting model produces 34 recommendations over roughly four weeks. Seventeen weeks in the first half go 10-7; the second half goes 9-8. The combined return on turnover is +0.65% — essentially flat.

What does that tell you about whether the model has an edge?

The honest answer is: almost nothing at all. Not “it is probably mediocre”. Not “the jury is out but leaning negative”. Genuinely nothing — the result is consistent with a strong model having a bad month and with a worthless model having an average one. Understanding why is the single most useful piece of statistical literacy a bettor can acquire, and it protects you in both directions: against abandoning something that works, and against trusting something that does not.

The arithmetic of noise

Betting results follow a binomial distribution. For a series of bets at a fixed win probability, the standard deviation of the number of wins is the square root of n × p × (1−p).

Take a genuinely strong model that wins 55% of its bets at standard -110 pricing. Over 34 bets, the expected number of wins is 18.7, and the standard deviation is the square root of 34 × 0.55 × 0.45 = 2.90 wins.

Two standard deviations either side covers roughly the middle 95% of outcomes. That means a 55% model will, in a normal month of 34 bets, plausibly return anywhere from about 13 wins to about 25 wins — a hit rate between roughly 38% and 73%.

Read that range again. A model with a real, substantial edge can produce a 38% month without anything being wrong. It can also produce a 73% month without being anywhere near that good. The month you happen to observe is close to uninformative.

How the range narrows

Bets	Expected wins at 55%	Roughly 95% of outcomes fall between
34	18.7	38% – 72%
100	55	45% – 65%
500	275	50.6% – 59.4%
1,000	550	51.9% – 58.1%
2,000	1,100	52.8% – 57.2%

Notice where the break-even line of 52.38% sits relative to those ranges. At 500 bets, the plausible band still includes outcomes below break-even. It is only somewhere past a thousand bets that a 55% model reliably separates itself from a coin flip at standard juice.

The formal version: to distinguish a 55% true rate from the 52.38% break-even with reasonable statistical confidence, you need on the order of 1,400 bets. At five bets a day, that is most of a year. At the pace of a single-sport service publishing a few picks a week, it is several years.

And if the true edge is smaller — 53.5% rather than 55% — the required sample grows into the tens of thousands.

Losing streaks are not evidence of anything

A 55% model loses 45% of the time. The probability of eight consecutive losses is 0.45 to the eighth power, or about 0.17%. That sounds reassuringly rare until you account for how many opportunities there are.

Across 500 bets, the expected number of eight-loss runs is roughly 0.84 — meaning a bettor placing 500 bets a year should expect to see one most years. Ten-loss runs occur less often but are far from exotic.

This is worth internalising before it happens rather than during. The emotional experience of eight straight losses is indistinguishable from the experience of a broken model. The statistics are not.

So what should you actually look at?

If results over any realistic period are too noisy to judge, you need lower-variance signals.

Closing line value. Whether you consistently obtain better prices than the market’s final number is a near-deterministic measurement rather than a coin flip. Patterns emerge in dozens of bets rather than thousands.

Calibration. If a model says 60% and those events happen close to 60% of the time across many predictions, the model is describing reality accurately. Calibration can be assessed across all predictions simultaneously, which is far more efficient than assessing profit.

Methodology. Whether the process is sound — proper out-of-sample testing, no lookahead bias, sensible feature selection — is something you can evaluate in an afternoon without waiting for a thousand bets.

Consistency of process rather than results. A source that publishes every pick, at the price available when published, including the bad ones, is telling you something a monthly profit figure cannot.

The uncomfortable implication for anyone selling picks

This cuts against every service in the industry, including ours. If a month cannot demonstrate an edge, then no service can honestly point at a month and claim one. Any advertised yield figure drawn from a short window is a statement about variance, not about skill — regardless of how good the number looks.

The correct posture for a subscriber is therefore: judge the process, track CLV, size your stakes so that a normal eight-loss run is survivable, and treat any claim built on a short record — good or bad — with the scepticism it deserves.

Frequently asked questions

How many bets do I need before my own results mean something?

Roughly 1,000 to 1,500 for a moderate edge at standard pricing, and considerably more if your edge is small. Below a few hundred bets, your win rate is dominated by variance.

Is a long losing streak a sign the model has stopped working?

Not by itself. Eight-loss runs are expected roughly once per 500 bets for a 55% model. A change in the underlying method, in the markets being targeted, or in closing line value is far more diagnostic than a streak.

Why do so many services advertise short-term returns then?

Because short windows produce dramatic numbers in both directions, and it is always possible to select a favourable one after the fact. A figure drawn from a few weeks is not evidence of an edge.

Does betting more per day speed this up?

It accumulates sample faster in calendar terms, but only if the additional bets carry the same edge. Adding lower-quality selections to reach a sample size faster reduces the average edge and defeats the purpose — and increases exposure.

Related reading

• Closing Line Value: The Only Early Signal That You Have an Edge — closing line value explained
• Backtesting vs Overfitting: How to Tell a Real Model From a Curve-Fitted One — spotting an overfitted betting model
• Kelly Criterion vs Flat Staking: How Much Should You Actually Bet? — Kelly criterion versus flat staking

Other news