Concept Primer · The Sports Page
The Sports Page
Making the numbers mean something since the first pitch
Concept No. 26Signal vs. Noise Back to The Sports Page →

Every Number Is Part Skill and Part Luck. The Trick Is Knowing How Much of Each.

No statistic is pure. Each one is a real signal — the thing you actually want to measure — buried in noise you don’t. The good news, and the whole of this idea: you can look up, for almost any number, exactly how much of it you need before the signal wins.
Tier 1 · The Two-Minute Version

A small number is almost all luck. A big one lets the skill show.

A hitter goes four-for-eight and his average reads .500. Nobody believes he is a .500 hitter, and rightly — eight at-bats is almost pure noise, the random luck of which balls found gloves. Let the season run to six hundred at-bats and the number settles into something real: the signal, his actual skill, has finally drowned out the luck. Every statistic works this way. It starts as mostly noise and, as the sample grows, slowly fills up with signal.

Here is the part that turns this from a truism into a tool. Different numbers fill up with signal at wildly different speeds. Some are trustworthy almost at once; others essentially never settle within the data you’ll ever have. A pitcher’s strikeout rate is mostly real after a few dozen batters. Whether hits happen to fall in against him barely becomes real across a whole career. Same box score, two numbers — one you can believe by April, one you should never quite trust.

Same amount of data, two very different numbers — one fills with signal fast, one crawls.
How much of the number is real, as the data piles up all skill half real (“stabilized”) all luck 0 a half-season a full season how much data you’ve collected → a “fast” stat (strikeout rate) a “slow” stat (the hits that fall in)

So the useful question about any number is never just “what does it say?” It is “how much of it is there, against how much it needs?” A big, impressive figure built on a thin slice of data is a rumour. The same figure, once the sample is deep enough, is a fact. Learn to ask which one you’re looking at, and you will stop being fooled by half the numbers people wave at you.

Tier 2 · If You Want to Go Deeper

Reliability, and the sample where a stat “stabilizes.”

Formally, what you observe is a sum: observed variation = true (skill) variation + error (luck) variation. The share that is real — the true part over the whole — is called reliability, and it runs from zero (all noise) to one (all signal). Collecting more data does not change a player’s true skill; it shrinks the error term, so reliability climbs toward one as the sample grows. That rising curve is the shape in the figure above.

Analysts pin this down with a stabilization point: the sample size at which a statistic correlates about 0.7 with a separate, equal sample of the same player — the moment roughly half its variation is finally skill. Those points are startlingly different from stat to stat. For pitchers, a strikeout rate stabilizes near seventy batters faced; a walk rate near a hundred and seventy; the batting average on balls in play only near two thousand, which is to say never within a season. The number of events a stat needs is a direct read on how much of it the player actually controls. Skills that are the player’s own — missing bats, throwing hard — firm up fast. Outcomes that the defense, the ballpark, and chance co-author firm up slowly, or not at all.

This is also the hinge between two ideas you already know. A number that stabilizes slowly is one whose early readings are mostly luck — so it will drift back toward the truth, which is regression to the mean. A number that stabilizes fast is one you can trust early — so it, and not the noisy one, is what you use to forecast. The reliable measure is the one that predicts; the unreliable one is the one that fools you into predicting from noise. Same coin, two faces — and the stabilization point tells you which face you’re holding.

Tier 3 · Take It to the Textbook

The full treatment: reliability, signal-to-noise, and shrinkage.

Where reliability comes from — the variance decomposition, the signal-to-noise ratio, split-half and test–retest correlation, Cronbach’s alpha, and the shrinkage (empirical-Bayes) estimators that pull an unreliable number back toward the mean by exactly the right amount — is developed in The Sports Page’s companion statistics textbook, a free, open, graduate-level text, with the R code to compute a stat’s reliability and find its stabilization point yourself. Read it free here — the same “read, play, learn” idea, one rung deeper.

Where this concept shows up in The Sports Page