← Blog · 📝 Article · 15 November 2026
Sample Size in Betting: How Many Bets Prove an Edge?
There is no single number of bets that proves a strategy works. The right sample depends on the size of the edge, the odds you bet at and how often you win. A rough guide for horse racing: a few hundred bets tell you whether a record is worth following, and thousands are needed before it is convincing. If you follow a service that publishes every selection before the race and keeps a permanent results history, such as our free and Pro lay signals at DonkeyRadar, the sample grows in public and you can judge it as it builds.
This guide explains the idea without heavy maths, shows a worked lay betting example and gives you a checklist you can apply to any record, ours included.
TL;DR
- Sample size is the number of bets behind a result. The smaller it is, the more luck can explain what you see.
- Uncertainty shrinks with the square root of the number of bets: to halve it you need four times as many bets.
- Small edges need huge samples. A peer-reviewed study of horse race betting put the bar for proving expertise at over 10,000 bets.
- A big count can still mislead when bets are linked or a few long-priced results dominate the profit.
- Judge a record by its count, its odds range, its largest results and whether every bet was published before the race.
Table of Contents
- What sample size means in betting
- Why small samples mislead
- How many bets does horse racing need?
- A worked example: sample size for a lay strategy
- Why a large sample can still mislead
- A checklist for reading any betting record
- Where DonkeyRadar fits
- Our view on sample size
- FAQ
- Sources
What sample size means in betting
Sample size is simply how many bets, races or observations sit behind a result. If you have placed 40 bets and are up 12 units, that is your sample. The question is whether 40 bets can tell you anything about the next 40.
Every betting result is made of two things: the real edge of the method (skill) and chance (variance). With few bets, chance is the louder of the two. As the count rises, chance averages out and the edge, if there is one, becomes visible. BetOwl’s UK guide puts the honest question this way: not “have I reached 500 bets?” but “how uncertain is my estimate, and could ordinary variation still explain what I see?”
Two terms are worth knowing:
- Strike rate is the share of your bets that win. For a lay bet, a win means the horse you laid loses.
- Standard error is the typical size of the gap between what you observe and the underlying truth. It is the measure of how much to trust a figure.
For a strike rate, the standard error is the square root of p × (1 − p) ÷ n, where p is the strike rate and n the number of bets. A rough 95% range is about two standard errors either side, as BetOwl explains.
Why small samples mislead
The table below shows the approximate 95% uncertainty around a 50% strike rate, using BetOwl’s figures.
| Number of bets | Approximate 95% range around a 50% strike rate |
|---|---|
| 100 | ±9.8 percentage points |
| 400 | ±4.9 |
| 1,000 | ±3.1 |
| 2,500 | ±2.0 |
| 10,000 | ±1.0 |
A 55% strike rate after 100 bets cannot be told apart from 50%. The range is wide enough to include both a losing method and a good one.
The pattern in the table is the one rule worth remembering: noise falls with the square root of the number of bets. Docteur Bet’s sample-size module states it plainly: the noise is halved only when the number of bets is multiplied by four. Its example is a +8% return over 150 bets at odds of 2.00, which still carries a 95% confidence interval from −8% to +24%. The true return could be negative, and it could be excellent. 150 bets cannot say which.
This is why a hot month proves very little and why a short, hand-picked period should never be treated as a track record.
How many bets does horse racing need?
The honest answer is more than most people expect. Three pieces of evidence help.
An academic benchmark. A peer-reviewed study of Australian off-track horse race betting, using 211 weekend races and Monte Carlo simulation, found that after allowing for the house advantage a gambler would need to place over 10,000 bets with net returns above 9% to be reasonably called an expert punter at the 5% significance level. Fewer bets would need even higher returns. That is a demanding standard, and it is meant to be: it stops lucky runs being mistaken for skill.
The smaller the edge, the bigger the sample. Tofiko’s guide shows that at even money, halving the edge quadruples the bets needed. Reaching two standard errors takes about 400 bets for a 10% edge, 1,600 for a 5% edge and 10,000 for a 2% edge. Longer prices cost more bets again, because a longer-priced bet swings results more.

The odds range matters. In Docteur Bet’s figures, at average odds of 2.00, detecting a true +5% edge at 95% confidence needs roughly 1,500 bets. At odds of 4.00 the same edge needs about 4,800.
Put together, a workable reading is:
| Bets in the record | What it tells you |
|---|---|
| Under 100 | Almost nothing. Treat as a hypothesis. |
| 100 to 300 | An early indication. Enough to decide whether to keep watching. |
| 300 to 1,000 | A fair picture of strike rate, a rough picture of profit. |
| 1,000 and over | The point where a large edge starts to separate from luck. |
| Several thousand | Needed to measure a modest edge with confidence. |
The bands are a guide built from the sources above, not a law. A larger edge needs fewer bets to show itself, and a thin one needs far more.
A worked example: sample size for a lay strategy
Lay betting turns the usual picture round. You back nothing; you act as the bookmaker on a horse you expect to lose. If it loses, you win the backer’s stake. If it wins, you pay out at the odds. (A lay bet on the Betfair Exchange is the type we cover in our Betfair lay betting strategy guide.)
That shape matters for sample size. Lay strategies win often and lose big. Here is a simple illustration with made-up round numbers, not our results:
- You lay at odds of 5.00 with a one-unit stake, so each win is +1 unit and each loss is −4 units.
- Break-even strike rate is therefore 80%: four wins offset one loss.
- Suppose the method truly wins 85% of the time. Expected profit is 0.85 × 1 − 0.15 × 4 = +0.25 units per bet, before commission.
How many bets would you need to see that edge clearly? Statometrics Academy gives the formula n = (t × σ ÷ edge)², where σ is the standard deviation of a single bet, edge is the expected return per unit and t is the target t-statistic, usually 2.
For our example, the standard deviation of one bet works out at about 1.79 units. Then n = (2 × 1.79 ÷ 0.25)², which is roughly 200 bets.
Now see what that means in practice:
| Bets | Rough 95% range for profit per bet (true value +0.25) |
|---|---|
| 100 | −0.10 to +0.60 units |
| 200 | 0.00 to +0.50 units |
| 400 | +0.08 to +0.43 units |
Three lessons come out of it.
- A high strike rate does not shorten the wait on its own. The losses are large and rare, so a stretch of 100 bets can contain two or three more losses than average and look bad, or two or three fewer and look brilliant.
- Strike rate and profit converge at different speeds. The strike rate can look steady while profit is still wide open, because one loss at 4 units cancels four wins.
- Edge size drives everything. If the true strike rate were 82% rather than 85%, the edge would be far thinner and the sample needed would climb sharply. Thin edges need the long record.
Our own claim is a strike rate of over 85% on lay signals across UK, Australian and US racing. Strike rate is only half the story at lay odds, so the odds each signal was priced at matter just as much, and our results history records both.
Why a large sample can still mislead
A big number is not automatically a good one. Four things to look for:
Linked bets. BetOwl points out that repeated selections may not be independent. In racing, selections exposed to the same going forecast or the same model version can share one error, so counting each as a fresh observation overstates the effective sample. Ten lays on one soft-ground card are closer to one observation than ten.
A few big results. BetOwl also notes that a win at 2.00 and one at 15.00 have very different effects, and a few long-priced winners may dominate a whole history. It recommends showing the spread of odds and the contribution of the largest wins, not one headline percentage. A simple test: remove the three biggest results and see whether the record still stands.
Slicing the data. Breaking a history into courses, months or price bands after the fact shrinks the real sample inside each slice. A claim that “handicaps at one track in spring” is profitable rests on far fewer bets than the total suggests.
Deciding afterwards. Statometrics warns against fixing the sample size after you have started, planning around a hoped-for edge rather than a realistic one, and counting linked bets as separate. Choose your test before the bets are placed.
There is also the mismatch between the sample you need and the sample you have. Tofiko’s arithmetic shows that at ten bets a week, proving a 2% edge could take around twenty years. That is why the volume of bets a method produces matters as much as its quality: a service producing signals every racing day builds a meaningful record far faster than one producing a handful a month.

A checklist for reading any betting record
Use this on any tipster, model or system, including your own notes.
- Count the bets. Fewer than 100 is a story, not evidence. Look for hundreds as a minimum and be wary of anyone showing only a short, hand-picked period.
- Check that every bet was recorded before the race. A record built from selections posted after the result is not a sample, it is a selection of winners.
- Look at the odds. Show the price each bet was taken or laid at, not just the result. Strike rate means little without it.
- Find the biggest results. Remove the top three and see what is left.
- Allow for the range. An observed 55% after 100 bets is a range of roughly 45% to 65%. Ask what the lower end would mean for your bank.
- Match staking to the losing run. More bets means longer likely losing runs, so stakes must be sized to survive them. Our lay betting system guide covers how staking tiers fit in.
- Look at yield, not only ROI. Our post on yield versus ROI explains why, and how to use closing-price evidence alongside it.
- Run an audit. Our eight-step tipster audit and our piece on spotting real edge with 500 bets walk through the process in detail.
For the maths behind prices, implied probability and the margin is worth reading first, because the margin is what every sample has to overcome.
Where DonkeyRadar fits
We publish every lay signal before the race and keep the results permanently available, so the sample you judge is complete, not curated. Selections come from statistical analysis of historical strike rates and live market prices rather than opinion.
The free plan gives you the daily lay signals. DonkeyRadar Pro is £29 per month after a seven-day trial and adds real-time alerts by email and Telegram, the full results history, API access for trading software and a dashboard with direct Betfair Exchange links and staking tier grading. If you want to see what a growing record looks like, start with today’s lay tips or go to DonkeyRadar and pick a plan.

If you are new to choosing what to lay, our checklist for picking horses to lay shows the process step by step. Please bet responsibly: gambling is for over 18s only, and you should only stake what you can afford to lose.
Our view on sample size
Most arguments about betting services are really arguments about sample size. Someone shows a good month, someone else calls it luck, and neither can be settled without a count.
That is why we publish every signal before the race and keep the results history permanent. A record that can be edited, trimmed or started late is not a sample. A record where every bet was visible beforehand, with its price, is one you can measure yourself.
We also think lay betting deserves its own reading of the numbers. A high strike rate is the headline, but the lay price decides what each loss costs, so the two have to be read together. Losses happen. It happens, and transparency is the whole point.
Our approach is built on tested statistical parameters rather than opinion, and the tooling around it, from staking tiers to API access, is there so the signals can be applied consistently over hundreds and thousands of bets. Consistency is what lets a sample mean something.
Donkey
FAQ
How many bets do I need to know if a strategy works?
There is no fixed number. A few hundred bets is a reasonable minimum for a first read, and thousands are needed for a modest edge. The smaller the edge and the longer the odds, the more bets you need.
Is 100 bets enough?
Not to prove anything. At a 50% strike rate, 100 bets leaves a 95% range of roughly ±9.8 percentage points, according to BetOwl. It is enough to decide whether a method is worth watching, not whether it is proven.
Does a high strike rate mean I need fewer bets?
Not necessarily. With lay betting the losses are larger than the wins, so profit can swing on a few results even when the strike rate looks steady. Read strike rate and odds together, and judge both over a long run.
Why does the sample have to be so large for horse racing?
Because racing results are noisy and edges are thin. The PubMed-indexed study found that over 10,000 bets with net returns above 9% were needed to show expertise at the 5% level. Noise only halves when you quadruple the bets.
Do bets on the same day count as separate samples?
Only partly. Bets exposed to the same ground conditions or the same model error are linked, and counting them as fully independent overstates how much you know. Treat linked bets with caution when you count.
Sources
- Sample Size in Betting Tests: No Magic Number
- Delusions of expertise: the high standard of proof needed to demonstrate skill at horserace handicapping
- Sample Size: How Many Bets Before You Know You Have an Edge
- Sample size, Docteur Bet
- How Many Bets to Prove an Edge? Sample Size for Betting