How many bets does it take to evaluate a performance?

The essentials in a few seconds

There is no universal number of bets beyond which a record becomes reliable. The amount of data needed to tell a method apart from plain variance depends on identifiable parameters — and that dependency can be calculated.

Why the question has no single answer

"How many bets?" invites a numeric answer in everyday language, when the real answer is a list of parameters. The same number of decisions can be plenty in one case and far too few in another, depending on what makes them up.

That's what sets this page apart from questions already covered elsewhere on the site: it doesn't ask which questions to put to a record, nor how to assess a tipster — it asks what the required amount of data mathematically depends on.

What the answer depends on

Six factors, all measurable on an existing record.

• The average odds the decisions are taken at — it directly determines the size of the variance.
• The associated win probability — the more extreme it is in either direction, the more concentrated or spread out the results are.
• The size of the assumed edge — a 1% edge and an 8% edge don't need the same volume, and by a wide margin.
• The variance produced by the two points above — odds and probability together determine how wide the possible swings around the expected value are.
• The spread of stakes — uneven stakes make an ROI harder to read than an equivalent volume of fixed stakes, because a handful of large-stake decisions can dominate the result.
• The metric observed — a strike rate, an ROI and a CLV don't converge on a stable value at the same speed.

A seventh, more subtle factor comes on top: correlation between bets. Several decisions tied to the same event, the same competition or the same reasoning count for less than their raw number suggests — they don't add as much independent information as the same number of genuinely distinct decisions.

Two profiles, two different volume needs

A small edge at short odds and a clear edge at long odds don't need the same number of decisions to become readable. A 1% edge at an average odds of 1.50 stays buried for a long time in a variance that's already small in size but insufficient to pick out such a small signal; an 8% edge at an average odds of 3.00 stands out much sooner, despite individually wider variance, because the gap to detect is itself far bigger.

In practice: the volume needed follows the gap between the signal you're looking for and the noise around it, not variance taken on its own. A clear signal breaks through faster than a faint one, even through more noise.

Canonical examples: 10 bets at +30%, 2,000 bets at +3%

Two records from the cluster, already cross-referenced on the ROI page, illustrate this mechanism directly. An ROI of +30% on 10 bets fits almost entirely within the range produced by variance on such a small sample — a few different results would have been enough to reverse it. An ROI of +3% on 2,000 bets has far less room to be explained by chance alone, without that guaranteeing it will repeat.

The second figure is smaller; it carries more information all the same, precisely because the volume behind it shrinks the share variance can occupy.

What the sample doesn't fix

A larger sample never makes an estimate more accurate. It only makes an existing gap more readable. If the process producing the decisions is structurally biased, more data only shows that bias more clearly — it doesn't correct it.

How to accumulate usable data

None of the above can be calculated without a record kept correctly, decision by decision. The tracking method — which columns to note, how to avoid gaps — is covered on a dedicated page.

The data to record for every bet →

Applying this to someone else

The same volume requirements apply to a tipster's or a forecaster's record. An example already worked out elsewhere on the site shows why a high strike rate on a sample made up of very short favourites proves nothing — it's not redone here, it's covered on the page about assessing a tipster.

How to assess a tipster's record →

The questions to ask any record, including your own →

No magic threshold appears on this page, nor anywhere else on this site: no one should promise one, and the number of decisions needed is calculated from the parameters above, never from a generic figure.

Dig into the market

Odds movements are only part of the story. Here are the next topics to read.

Frequently asked questions