Sample size: the number next to the number

Every rate on this site is printed with the count it came from, and that is not a house style. A percentage on its own hides the one thing that decides whether it means anything, which is how many games went into it.

What a fair coin does at five, ten and twenty games

Start with a player whose true rate is exactly 50% — a genuine coin flip against the number on offer. Watched five times, that player clears the number somewhere between once and four times in 30 of every 32 runs. That is a record anywhere from 20% to 80%, produced by luck alone, with nothing about the player changing at all.

Ten games narrows it hardly at all: 2 to 8 clearances, the same 20% to 80% band, in 1,002 of every 1,024 runs. Twenty games is the first sample that rules much out, and it still leaves 6 to 14 clearances — 30% to 70% — in 1,005,176 of every 1,048,576.

Now set that against what a price is claiming. A price of 1.25 implies an 80% chance; 5.00 implies 20%. At five games and at ten, the range a genuine coin flip covers on its own is that entire span. A five-game record cannot tell apart a market the book has at 80% from one it has at 20%, because both readings sit comfortably inside what an even chance produces by itself. This is the whole reason a rate without its count is not yet information.

The four floors this site uses

None of these four is a style rule. Each is a named constant in the code, and each decides something a reader can see: whether a chip appears at all, whether a badge is printed beside a rate, whether a page is offered to a search engine. They sit at three, four and five games, which is low — the figure above is the argument that even twenty is not many — and they sit there because below them the site would be publishing arithmetic it cannot stand behind.

Three games

A split bucket on a player's props card — Home, Away, Wet, Dry, the head-to-head, at-venue, Wins, Losses — is hidden below three games in the window on screen. The explanation doors beside those chips, the ones labelled “what this says”, need three games on both sides before they appear at all: a comparison where one of the two piles holds two games is not a comparison. Open any player from the AFL board and look at the Splits row.

Four games, on the props card and in the projection

Below four games the hit rate wears a thin badge, and the share text spells the reason out rather than shrinking the font. Below four games there is no projection at all. An adjustment resting on fewer than four games is printed as too thin to count and contributes exactly zero — shown rather than dropped, because an adjustment that quietly vanished would leave a reader wondering whether it had been counted.

Four games again, on the two islands

The venue island and the wet-versus-dry island under the Matchup Analyser heading carry their own footnote at the same floor, naming the count and calling it thin evidence. The number is the same and the code is separate, because those islands read a player's whole multi-season record while the props card reads the window the reader chose, and the two counts are different facts.

Five games

Below five, a player page is left out of the sitemap and asks search engines not to index it, and the compare board holds to the same floor before it will print a percentage instead of a raw fraction. Being absent from a search result is a real cost, and this site would rather pay it than be found saying something off four games.

“Could that just be luck?”

Every split door carries a button with exactly that wording, and what it opens is not a significance test. It is a count: how many results in the smaller of the two groups would have had to land the other way for both groups to read the same. Two games and the split disappears is a sentence a reader can check against the bars already on the screen. A p-value is a number taken on trust, and taking a number on trust is the habit this whole site is arranged against.

The window chips are a sample-size control

L5, L10, L20 and Season sit above the game log, and they are usually read as a recency control. They are also the denominator. A player page opens on the last twenty games: long enough that one game cannot swing the rate ten points, short enough to still describe this season. A player with no games this season opens on the full record instead, and the page says so rather than showing an empty chart. L5 is a form check rather than a rate — five games is the top row of the figure above, and the top row spans 20% to 80%.

How to read a number with its sample

Read the count before the percentage. Ask what band that count alone would allow. If the price sits inside that band, the honest reading is that the sample cannot separate the two, and that is a real answer rather than a missing one. It is also the most common answer on any given market, which is why this site prints the working instead of a shortlist.