Definition
A z-score is the number of standard deviations an observation lies above or below the mean of its reference distribution. It is computed by subtracting the mean and dividing by the standard deviation. The transformation re-centres the distribution at zero and rescales it to unit variance, so any observation, regardless of its original units, lives on the same standardised axis.
The construct is foundational because it solves two problems at once: it makes incomparable quantities comparable, and it expresses how unusual any single observation is relative to the rest of its distribution. A z of one is ordinary; a z of three is rare; a z of five is, under normality, vanishingly so.
Why it matters
How it works
Calculation requires two parameters — the population mean and standard deviation — and one observation. In practice both parameters are estimated from a sample, and the choice of sample window matters enormously. A z-score computed against a full-history mean is appropriate for stationary processes; against a rolling window it becomes a time-localised measure that adapts to regime shifts, at the cost of being noisier and reflecting only recent context. Many quantitative strategies build their entire signal generator on rolling z-scores of price spreads, returns, or volume.
Two cautions sit on top of the formula. Real-world distributions are rarely normal — financial returns in particular have fat tails — so a z of four does not imply the one-in-thirty-thousand frequency the normal distribution would suggest. Reading z-scores through a heuristic of "look hard at anything beyond three sigma" is sound; treating them as exact tail probabilities is not. And the mean and standard deviation used for normalisation must be calculated using only data available at the observation time; using future data inflates apparent signal and is a common form of look-ahead bias in time-series research.